A park fire linkage execution sequence control method

Abstract

The application relates to the technical field of fire monitoring, and discloses a park fire linkage execution sequence control method, which comprises the following steps: receiving events reported by wireless networking in a park and converting the events into uniform event vectors, and merging the events into event clusters based on the correlation between the events; extracting multi-dimensional characteristic quantities, calculating evidence consistency scores and real risk scores; generating a candidate linkage action set for each event cluster, constructing an action execution sequence diagram, solving a priority execution sequence, and obtaining a total execution queue by summarizing; outputting the first batch of candidate linkage actions and issuing them, updating the evidence consistency scores and the real risk scores, obtaining an updated total execution queue, and executing subsequent candidate linkage actions; and reversely correcting early warning parameters according to a closed-loop conclusion to generate early warning information; the application unifies and analyzes concurrent events from different building units, different fire subareas and different types of nodes in the same period, thereby improving the efficiency of park fire linkage disposal and the rationality of resource allocation.

Description

Technical Field

[0001] This application relates to the field of fire monitoring technology, and more specifically, it relates to a method for controlling the execution sequence of fire-fighting linkage in a park. Background Technology

[0002] With the continuous development of industrial parks, factories within factories, and multi-entity scenarios, park fire protection systems are no longer limited to alarm uploads triggered by a single detector. Instead, they have gradually evolved into a complex, interconnected system involving smoke detectors, heat detectors, manual alarms, input / output controls, video surveillance, on-site response terminals, and management platforms. In practical park applications, fire management often needs to simultaneously receive alarm, fault, recovery, manual feedback, and verification information from different building units, functional zones, and different types of nodes. Especially under wireless network access conditions, the concurrency of multi-source events is even stronger, causing park fire response to shift from a single-point alarm response problem to a multi-event collaborative response problem.

[0003] Meanwhile, to reduce cabling costs and improve the flexibility of upgrades, an increasing number of fire sensing and linkage nodes in industrial parks are adopting Z-Wave or other wireless mesh networks to access the fire management system. While wireless networking offers advantages such as flexible deployment and convenient expansion, it can also lead to issues like link quality fluctuations, routing delays, node offline reconnection, duplicate message reporting, and inconsistent arrival times of different information. Consequently, within the same time period, the control platform may receive multiple event messages from different areas, nodes, and categories with varying degrees of reliability, making it unsuitable to use a fixed process for determining the order of linkage actions.

[0004] A Chinese patent with authorization announcement number CN116665424B discloses an automatic fire alarm and fire-fighting linkage system. It includes a first data acquisition module that collects environmental data of power lines in the monitored area; a second data acquisition module that collects power line information in the monitored area; a first data processing module that analyzes and processes the environmental data to generate an environmental impact coefficient corresponding to the monitored area, compares the environmental impact coefficient with a preset gradient threshold, and generates an environmental impact marker; a second data processing module that receives power line information, analyzes and processes the power line information, generates a fire symptom impact coefficient corresponding to the power line, compares the fire symptom impact coefficient with a preset fire symptom impact coefficient gradient reference value, and generates a fire symptom marker; and an early warning generation module that generates fire early warning information for the corresponding monitored areas based on the environmental impact markers and fire symptom markers for each monitored area.

[0005] However, when multiple wireless nodes report different types of events such as abnormal smoke, abnormal temperature rise, equipment failure, communication recovery, and manual triggering at similar times, existing technologies struggle to comprehensively consider the correlation between events, the consistency of evidence, regional risk differences, and changes in feedback information to uniformly prioritize and dynamically adjust multiple candidate coordinated actions. This can easily lead to situations such as high-risk events not being prioritized, unreasonable order of review and notification actions, resources being occupied by low-priority events, and the inability to promptly correct the execution order based on new feedback, thus affecting the efficiency and reliability of fire-fighting coordinated response in the park.

[0006] Therefore, existing technologies lack a method for controlling the execution sequence of fire-fighting linkage in scenarios with concurrent multi-source events in wireless networking within a park, in order to solve the problem of the difficulty in reasonably determining the execution sequence of linkage actions under concurrent multi-source events. Summary of the Invention

[0007] The purpose of this application is to address the problem in existing technologies that, under the condition of multiple concurrent events in a park's wireless network, it is difficult to comprehensively consider the correlation between events, the degree of consistency of evidence, regional risk differences, and changes in feedback information, resulting in the difficulty in reasonably determining the execution sequence of fire-fighting linkage actions and insufficient dynamic adjustment capabilities. This application provides a method for controlling the execution sequence of fire-fighting linkage actions in a park.

[0008] This application provides a method for controlling the execution sequence of fire alarm linkage in a park, including:

[0009] Within a preset sliding time window, events reported by the wireless network within the park are received, and the events are converted into unified event vectors. Based on the correlation between events, the unified event vectors are grouped into one or more event clusters.

[0010] For each event cluster, multidimensional features are extracted, and an evidence consistency score is calculated based on the multidimensional features. The true risk score is then calculated by combining the preset regional basic risk coefficient.

[0011] Based on the evidence consistency score and the true risk score, a set of candidate linkage actions is generated for each event cluster. An action execution sequence diagram is constructed based on the set of candidate linkage actions, and the priority execution sequence is solved. The total execution queue is obtained by summarizing the priority execution sequence based on the action execution sequence diagram.

[0012] The first batch of candidate linkage actions are output and issued according to the total execution queue. Within the preset listening time, the evidence consistency score and the real risk score are updated according to the feedback information. The remaining candidate linkage actions are partially rearranged to obtain the updated total execution queue. Subsequent candidate linkage actions are executed according to the updated total execution queue.

[0013] After the event cluster completes the disposal closed-loop, the early warning parameters are corrected backward according to the closed-loop conclusion, and early warning information is generated based on the early warning parameters.

[0014] In the above technical solution, the technical effects and advantages provided by this application are as follows:

[0015] This application receives multi-source events in the园区 wireless network within a preset sliding time window, converts each event into a unified event vector, and then constructs an event cluster based on the association relationship between events. Thus, it can uniformly merge and comprehensively analyze concurrent events from different building units, different fire protection zones, and different types of nodes within the same period, rather than only processing single alarm signals separately. This can effectively adapt to the actual scenarios of inconsistent arrival order of messages, repeated reporting, insertion of recovery information, and concurrent occurrence of multiple types of events in the园区 wireless network environment, and improve the integrity and accuracy of multi-source event recognition and processing.

[0016] This application calculates the evidence consistency score by extracting the multi-dimensional feature quantities of the event cluster, and calculates the real risk score by combining the regional basic risk coefficient, the influence coefficient of key objects, and the event intensity aggregation value. This makes the linkage control no longer rely on a single alarm trigger, but can comprehensively reflect the consistency degree between multi-source evidence and the risk differences in the area where the event is located. Thus, it can more accurately distinguish high-trust and high-risk events from low-trust events caused by node anomalies, link fluctuations, or local false alarms, and improve the reliability of园区 fire linkage judgment.

[0017] This application generates a corresponding candidate linkage action set for each event cluster based on the real risk score and the evidence consistency score, and further constructs a candidate linkage action execution sequence diagram. On the premise of satisfying constraint conditions such as forward and backward dependence relationships and mutual exclusion relationships, the priorities of the candidate linkage actions are solved and a total execution queue is formed. Thus, it realizes the unified sorting control of multiple event clusters and multiple types of linkage actions, enables candidate linkage actions with high risk, high time efficiency value, and high evidence clarification value to be executed first, and solves the problem that it is difficult to reasonably determine the execution order of linkage actions under the condition of concurrent multiple events in the prior art, thereby improving the efficiency of园区 fire linkage disposal and the rationality of resource allocation.

[0018] After the first batch of candidate linkage actions are issued, this application can continuously monitor the feedback information within the preset monitoring duration, map the new feedback information to incremental evidence, recalculate the evidence consistency score and the real risk score of the event cluster, and locally rearrange the unexecuted candidate linkage actions and update the total execution queue on the premise of keeping the existing forward and backward dependence relationships and mutual exclusion relationships unchanged. This makes the linkage execution order can be dynamically corrected with the emergence of new information such as video review results, manual feedback, continuous alarms, and fault recovery, avoiding the execution order being fixed once triggered, and enhancing the real-time adaptability and execution reliability in the process of园区 fire linkage disposal.

[0019] This application modifies the early warning parameters based on the closed-loop conclusion and generates corresponding early warning information, thereby forming a complete control link with organic connection between event reception, risk assessment, action sequencing, feedback reordering and closed-loop early warning. This not only improves the linkage and handling effect in the current concurrent event scenario, but also facilitates the continuous optimization of early warning strategies and sequencing criteria in subsequent similar event scenarios. Attached Figure Description

[0020] Figure 1 This is a module example diagram of a method for controlling the execution sequence of fire-fighting linkage in a park, as described in this application;

[0021] Figure 2 This is an example diagram of the overall execution queue obtained from a method for controlling the execution sequence of fire-fighting linkage in a park, as described in this application.

[0022] Figure 3 This is an example diagram illustrating the execution of subsequent candidate linkage actions in a park fire linkage execution sequence control method according to this application. Detailed Implementation

[0023] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0024] A method for controlling the execution sequence of fire-fighting linkage in a park is applied to a park fire management environment. The park has multiple building units, factory-within-a-factory units, or functional zones. Each zone is equipped with smoke detectors, heat detectors, manual alarm nodes, input / output control nodes, audible and visual alarm nodes, video surveillance nodes, and gateway nodes. At least some nodes are connected to the park fire control platform via Z-Wave or other wireless mesh networks. The park fire control platform is communicatively connected to at least one linkage control host, at least one video access unit, at least one personnel message push terminal, and at least one on-site handling terminal. It allows the reception of multiple event inputs from different sources, with different content, and different reliability within the same time window, and allows the reception of feedback information during the execution of linkage actions to trigger sequence reordering.

[0025] The method is as follows Figure 1 As shown, it includes:

[0026] Step 100: Receive events reported by the wireless network within the park within a preset sliding time window, convert the events into unified event vectors, and group the unified event vectors into one or more event clusters based on the correlation between events.

[0027] Within a preset sliding time window, all newly arriving events in the campus wireless network are received. The preset sliding time window refers to a continuous time interval with the current time as the endpoint and a preset time length as the span. This time interval slides synchronously as the current time progresses, ensuring that the system always focuses on events arriving in the most recent period. The preset time length is configured by the system administrator based on the event reporting frequency and network transmission latency of the campus wireless network. In campuses with high node density and frequent event reporting, it can be appropriately shortened to improve response sensitivity, while in campuses with low node density, it can be appropriately extended to ensure sufficient collection of related events. The events originate from wireless nodes in wireless networks with different protocols, device types, and reporting formats, including but not limited to Z-Wave protocol nodes, Zigbee protocol nodes, and private wireless protocol nodes.

[0028] The reason for using a preset sliding time window instead of processing events in real time is that in a campus wireless network, the same risk source often triggers responses from multiple nodes within a short period of time. These response events arrive at different times due to wireless transmission delays and protocol differences. If the system triggers linkage actions for each event one by one, it may make one-sided linkage decisions before enough related events have been collected, resulting in an unreasonable priority of linkage actions. By setting a preset sliding time window, the system can accumulate a sufficient number of related events within a reasonable time span, providing a sufficient data foundation for subsequent multi-dimensional correlation calculations and event cluster construction. The sliding time window design ensures that the system neither delays responses due to excessively long waiting times nor misses related events due to excessively short collection times, achieving a balance between response timeliness and event completeness.

[0029] Each event is converted into a unified event vector; the unified event vector includes the event timestamp, node identifier, region identifier, event category, event intensity, set of network quality parameters, and original payload summary.

[0030] The reason for converting heterogeneous events into a unified event vector is that events reported by nodes of different protocols in the campus wireless network differ in data format, field definition, and encoding method. If the system directly performs correlation calculations on the original heterogeneous events, it is necessary to design a separate correlation algorithm for each protocol combination. The algorithm complexity increases exponentially with the increase of protocol types and is difficult to maintain and expand. By converting all heterogeneous events into a standardized vector format that includes event timestamps, node identifiers, region identifiers, event categories, event strengths, network quality parameter sets, and original payload summaries, the system can perform undifferentiated correlation calculations on events from different protocols in a unified data space. This allows subsequent temporal proximity functions, spatial proximity functions, event category correlation functions, and network feature similarity functions to be operated on the same data structure, significantly reducing the complexity of algorithm design and system implementation. The reason for selecting the above seven dimensions as input features in the unified event vector is that these seven dimensions comprehensively characterize events from the perspectives of time, space, semantics, strength, network quality, and original data, providing sufficient and complementary information input for subsequent multidimensional correlation calculations.

[0031] The event timestamp is a unified time after clock correction, with an accuracy to the millisecond level. The clock correction method involves each wireless node carrying its local timestamp when reporting an event. When the event is forwarded to the park's fire control platform via the gateway node, the gateway node records the gateway reception timestamp. The park's fire control platform calculates the total clock offset of the wireless node based on the gateway reception timestamp and the node's local timestamp, and adds the total clock offset to the node's local timestamp to obtain the corrected event timestamp. The total clock offset is calculated by the park's fire control platform periodically sending clock synchronization commands to each gateway node. Upon receiving the clock synchronization command, each gateway node records the difference between the platform's sending time and its local time as the gateway clock offset. When forwarding wireless node events, the gateway node records the difference between the node's local timestamp and the gateway's received timestamp as the node's clock offset relative to the gateway. The sum of these two offsets is the total clock offset of the node relative to the platform. The clock synchronization command issuance cycle is configured by the system administrator based on the node's clock drift rate. In node environments with large clock drift, this cycle can be appropriately shortened to improve correction accuracy. For nodes that do not support carrying local timestamps, the system directly uses the gateway's received timestamp corrected by the gateway clock offset as the event timestamp. Through the above two-level clock correction mechanism, the event timestamps of nodes under different protocols and gateways are unified to the time base of the park's fire control platform, ensuring that the calculation results of the subsequent time proximity function have physical meaning.

[0032] The reason for adopting a two-level clock correction mechanism is that wireless nodes of different protocols in the campus wireless network each maintain independent local clocks. Due to differences in crystal oscillator accuracy and accumulated drift, these local clocks have a non-negligible deviation from the time reference of the campus fire control platform. If the uncorrected node local timestamps are used directly for subsequent time proximity function calculations, two events that actually occur almost simultaneously may be calculated as events with a large time interval due to clock deviations, resulting in distorted correlation calculation results and affecting the accuracy of event cluster construction. By adopting a two-level offset superposition method from the platform to the gateway and from the gateway to the wireless node, clock correction can be completed through the gateway node as an intermediate bridge without requiring all nodes to directly synchronize with the platform. This hierarchical correction method reduces the requirements for node communication capabilities while ensuring correction accuracy, making it suitable for campus environments with multi-protocol hybrid networks.

[0033] The node identifier is a unique code for the wireless node that reported the event; the area identifier is the fire zone code corresponding to the wireless node in the park spatial mapping information, which includes the correspondence between nodes and buildings, floors, sub-factories, fire zones, evacuation routes, and key equipment areas; the event category range includes alarm, fault, recovery, node offline, manual trigger, and inspection anomaly; the event intensity is a quantified value of the severity of the event under its category, ranging from 0 to 1, where 0 represents the lowest intensity and 1 represents the highest intensity; the specific value of the event intensity is obtained by normalizing the deviation of each type of wireless node from its sensor readings to the preset alarm threshold, for example, smoke detectors are normalized based on the ratio of smoke concentration to the alarm concentration threshold, and heat detectors are normalized based on the ratio of temperature to the alarm temperature threshold; the network quality parameter set includes the network routing hop count, response status, retransmission count, and signal quality indicators at the time the event was reported; the original payload summary is a summary of key fields in the original data frame of the event.

[0034] This step does not directly trigger the linkage action based on a single event. Instead, it merges all unified event vectors within the preset sliding time window into the park-level unified processing queue. This queue includes all unified event vectors arriving within the current sliding time window, and the total number of events in the queue is recorded as the total number of events in the current window.

[0035] The reason for adopting a unified processing queue at the park level instead of processing separately by region or protocol is that in a park fire scenario, the same risk source may simultaneously affect multiple types of wireless nodes in multiple fire zones. If separate processing queues are established by region or protocol, cross-regional and cross-protocol related events cannot be effectively identified and merged, causing the system's judgment of risk sources to be limited to the perspective of a single region or single protocol. By merging all unified event vectors into the same park-level queue, the system can perform event correlation calculations globally, discover cross-regional and cross-protocol event correlations, and thus more accurately identify the real risk source and generate a sequence of linked actions for it.

[0036] After forming a unified processing queue, the association calculation is performed on multiple unified event vectors in the queue, and events that meet the association conditions are merged into the same event cluster. The event cluster refers to a set of events that are related in time, space or semantics, and these events may correspond to the same real risk source.

[0037] Specifically, the correlation degree between any two unified event vectors in the unified processing queue is calculated; the correlation degree is equal to the sum of four items: the product of the time proximity function value and the preset time proximity weight coefficient, the product of the spatial proximity function value and the preset spatial proximity weight coefficient, the product of the event category correlation function value and the preset event category weight coefficient, and the product of the network feature similarity function value and the preset network feature weight coefficient.

[0038] The reason for using a weighted summation method based on four dimensions—time, space, event category, and network features—to calculate the correlation degree is that determining whether two events originate from the same risk source in a park fire safety scenario requires a comprehensive evaluation from multiple independent perspectives. Relying solely on the time dimension would misclassify unrelated events from different areas within the same time window as the same risk source; relying solely on the spatial dimension would misclassify unrelated events from different time periods within the same area as the same risk source; and relying solely on event category would fail to distinguish the differences between similar events in different areas. By multiplying the correlation function values ​​of the four dimensions by their corresponding weight coefficients and then summing them, the system can comprehensively consider information from four aspects: temporal proximity, spatial proximity, semantic coupling, and network feature similarity. This ensures that the correlation degree calculation result fully reflects the actual degree of correlation between the two events. The weighted summation method allows for flexible adjustment of the contribution of each dimension through weight coefficients, adapting to the differences in the importance of each dimension in different park scenarios, while ensuring that the correlation degree value always falls between 0 and 1, facilitating comparison with a preset merging threshold.

[0039] The sum of the four weight coefficients—preset time proximity weight coefficient, preset spatial proximity weight coefficient, preset event category weight coefficient, and preset network feature weight coefficient—is equal to 1, and each weight coefficient is greater than 0. The values ​​of the above four weight coefficients are all greater than 0 and less than 1. The specific values ​​are configured by the system administrator according to the actual scenario of the park. In parks with dense buildings and compact partitions, the preset spatial proximity weight coefficient can be appropriately increased. In parks with complex wireless network environments and frequent interference, the preset network feature weight coefficient can be appropriately increased. When adjusting, it is necessary to ensure that the sum of the four weight coefficients is always equal to 1.

[0040] The time proximity function is used to measure the proximity of two events in the time dimension. It is calculated by taking the natural constant as the base and the negative value of the quotient obtained by dividing the absolute value of the difference between the timestamps of the two events by a preset time decay constant as the exponent, and then obtaining the exponential function value.

[0041] The reason for using the negative exponential decay function as the time proximity function is that this function has the mathematical properties of being monotonically decreasing, continuously differentiable, and having a range strictly between 0 and 1. It can naturally map the time interval between two events into a normalized value representing the degree of proximity. When two events occur almost simultaneously, the time interval approaches 0, and the negative exponential function value approaches 1, indicating that the two are highly close in the time dimension. When the time interval between two events is much larger than the preset time decay constant, the negative exponential function value approaches 0, indicating that the correlation between the two in the time dimension can be ignored. Compared with the linear decay function, the negative exponential decay function decays more slowly in the region with a small time interval and decays more quickly in the region with a large time interval. This nonlinear decay characteristic is more in line with the physical law of fire event propagation, that is, multiple node responses triggered by the same risk source are usually concentrated in a short time window, and the correlation probability drops sharply after the time window is exceeded. The preset time decay constant is a key parameter for controlling the decay rate, and its value directly determines the sensitivity of the system to time correlation.

[0042] The preset time decay constant is configured by the system administrator based on the typical event propagation delay of the campus wireless network. The value range is a positive real number greater than 0. It can be appropriately reduced in environments with fast node response speed and low network latency, and appropriately increased in environments with high network latency. The absolute value of the difference between two event timestamps represents the time interval between the two events. The smaller the time interval between two events, the closer the value of the time proximity function is to 1. When the time interval is much larger than the preset time decay constant, the value of the time proximity function approaches 0.

[0043] The spatial proximity function is used to measure the proximity of two event regions in the park's spatial mapping. Its value is determined as follows: if two events have the same region identifier, the spatial proximity function value is 1; if two events have different region identifiers but are adjacent in the park's spatial mapping, the spatial proximity function value is a preset adjacent region spatial correlation coefficient. This preset adjacent region spatial correlation coefficient ranges from greater than 0 to less than 1. This value is configured by the system administrator based on the density of the park's building structure. It can be appropriately increased when adjacent zones are physically close and there are no fire barriers, and appropriately decreased when there are firewalls or large physical distances between adjacent zones; if two events have neither the same nor adjacent region identifiers, the spatial proximity function value is 0. Adjacent means that two regions share a physical boundary or belong to different floors of the same building unit in the park's spatial mapping information.

[0044] The event category correlation function is used to measure the semantic coupling degree between two events. It is calculated by querying a preset event category coupling matrix to obtain the semantic coupling degree corresponding to each event category. The preset event category coupling matrix is ​​a symmetric square matrix, with its rows and columns corresponding to six event categories: alarm, fault, recovery, node offline, manual trigger, and inspection anomaly. Each element in the matrix has a value between 0 and 1, representing the semantic coupling degree between the two event categories. The preset event category coupling matrix is ​​configured by the system administrator based on park fire protection business experience and historical event correlation analysis results. The semantic coupling degree of the same event category with itself is 1. When there is a strong coupling relationship between two event categories, the corresponding semantic coupling degree is close to 1; when the coupling relationship between two event categories is weak, the corresponding semantic coupling degree is close to 0.

[0045] The network feature similarity function is used to measure the similarity between two events in network transmission characteristics. It is calculated by dividing 1 by 1 and adding the sum of the normalized Euclidean distance between the network quality parameter sets of the two events, and taking the quotient as the network feature similarity function value.

[0046] The aforementioned temporal proximity function values, spatial proximity function values, event category correlation function values, and network feature similarity function values ​​are all normalized correlation components output by the corresponding functions, used to characterize the strength of the correlation between two unified event vectors from different dimensions. Specifically, the temporal proximity function value characterizes the degree of synchronization between two events in their corrected occurrence time; a larger value indicates that the two events are more likely to be responses generated by the same risk source at similar times. The spatial proximity function value characterizes the degree of spatial connectivity or adjacency of the fire protection zones to which the two events belong in the park's spatial mapping; a larger value indicates that the two events are more likely to be affected by the same spatial risk source. The event category correlation function value characterizes the strength of the accompanying, causal, or mutually corroborating relationship between two event categories in fire protection business semantics; a larger value indicates that the two event categories are more likely to point to the same risk process. The network feature similarity function value characterizes the similarity between two events in wireless network transmission paths, link quality, and communication states; a larger value indicates that the two events are more likely to originate from the same risk response process under similar network topologies or communication environments.

[0047] The reason for introducing a network feature similarity function is that multiple wireless nodes triggered by the same risk source in a campus wireless network are usually located in similar network topology positions, and their network routing hop count, retransmission count, and signal quality indicators at the time of event reporting often exhibit similar characteristic patterns. When a real fire occurs in a certain area, multiple wireless nodes in that area usually have small differences in their network quality parameter sets because they share similar wireless transmission paths and network environments. Conversely, if two events are close in time and space but have large differences in their network characteristics, it may indicate that the two events come from different network branches or different physical environments, and their correlation is different. The similarity needs to be appropriately reduced; using the inverse transformation of the normalized Euclidean distance as the network feature similarity function can map the distance between multidimensional network quality parameters to a similarity value between 0 and 1. The smaller the distance, the higher the similarity, and the larger the distance, the lower the similarity, but it is always greater than 0, avoiding the extreme case of negative or zero similarity due to large differences in network features; the purpose of normalizing each component in the set of network quality parameters is to eliminate the differences in the dimensions and numerical ranges between different components, so that each component has an equal contribution weight in the Euclidean distance calculation, and avoids the components with larger numerical ranges dominating the distance calculation results.

[0048] The normalized Euclidean distance is calculated as follows: First, each component in the network quality parameter set of each of the two events is normalized. This is done by subtracting the minimum value of each component from all events, dividing by the sum of the difference between the maximum and minimum values ​​of the component in all events plus a preset minimum positive number, so that the value range of each component is mapped to between 0 and 1. Then, the Euclidean distance between the two normalized network quality parameter sets is calculated and divided by the arithmetic square root of the number of components in the network quality parameter set to obtain the normalized Euclidean distance. The preset minimum positive number is a fixed positive number that approaches zero to prevent the denominator from being zero, and it does not need to be adjusted according to the scenario. Through the above calculation method, when the network features of two events are exactly the same, the normalized Euclidean distance is 0, and the value of the network feature similarity function is equal to 1. When the difference in the network features of two events increases, the normalized Euclidean distance increases, and the value of the network feature similarity function monotonically decreases and approaches 0 but is always greater than 0. This ensures that the value range of the network feature similarity function always falls within the interval of greater than 0 and less than or equal to 1, and there will be no negative values ​​or values ​​exceeding 1.

[0049] A preset merging threshold is set. When the correlation between two unified event vectors is greater than or equal to the preset merging threshold, the two events are grouped into the same event cluster. The preset merging threshold ranges from 0 to 1. This value is configured by the system administrator according to the strictness of event correlation in the park. In scenarios where stricter differentiation of different risk sources is required, it can be appropriately increased to reduce erroneous merging. In scenarios where broader aggregation of related events is required, it can be appropriately decreased to increase the coverage of event clusters. After calculating the correlation between all events in the unified processing queue pairwise, the transitive closure method is used to complete the initial construction of event clusters. If the first event and the second event are grouped into the same event cluster, and the second event and the third event are grouped into the same event cluster, then the first event and the third event are also grouped into the same event cluster. The transitive closure refers to the transitive extension of the correlation relationship, so that all directly or indirectly related events are grouped into the same event cluster.

[0050] The reason for using the transitive closure approach to construct event clusters is that the correlation between multiple events triggered by the same risk source in a park fire scenario may exhibit a chain-like transmission characteristic. For example, the correlation between smoke detector node A and heat detector node B exceeds the preset merging threshold, and the correlation between heat detector node B and manual alarm node C also exceeds the preset merging threshold. However, the correlation between smoke detector node A and manual alarm node C does not reach the preset merging threshold due to the large difference in event categories. In this case, if the transitive closure approach is not used, smoke detector node A and manual alarm node C will be assigned to different event clusters, causing the system to split the response events of the same risk source into multiple independent event clusters for separate processing, reducing the accuracy of risk assessment and the coordination of linkage actions. Through the transitive closure approach, the system can group all directly or indirectly related events into the same event cluster, ensuring that all response events of the same risk source are treated as a whole for subsequent evidence consistency scoring and true risk scoring calculation.

[0051] After the transitive closure merging is completed, an intra-cluster consistency check is performed on each initially constructed event cluster to prevent unrelated events from being over-merged into the same event cluster due to the chain bridging effect. The intra-cluster consistency check includes the maximum intra-cluster spatial span constraint, the intra-cluster central correlation constraint, and the intra-cluster minimum event category consistency constraint. The maximum intra-cluster spatial span constraint involves calculating the number of different fire zones covered by the area identifiers of all events in the event cluster. If this number exceeds a preset maximum spatial span threshold, the event cluster is determined to be over-merged and needs to be split. The preset maximum spatial span threshold is a positive integer greater than or equal to 2. This value is configured by the system administrator according to the granularity of the fire zone division and the building layout of the park. It can be appropriately increased in parks with small and densely arranged zones, and appropriately decreased in parks with large zones. The intra-cluster central correlation constraint involves calculating the correlation between each event in the event cluster and the central event of the event cluster. The central event is the event with the largest sum of correlations with all other events in the event cluster. If the correlation between any event and the central event is less than the preset maximum spatial span threshold, the event cluster is considered over-merged and needs to be split. If a lower threshold for central correlation is set, the event is deemed to have insufficient correlation with the center of the event cluster and needs to be removed from the current event cluster. The preset lower threshold for central correlation is greater than 0 and less than a preset merging threshold. This value is configured by the system administrator based on the strictness of consistency within the event cluster and is usually set as a proportional coefficient less than 1 times the preset merging threshold. The minimum event category consistency constraint within the cluster includes calculating the proportion of the number of events pointing to the same risk type in the event cluster to the total number of events in the event cluster. If this proportion is less than the preset minimum category consistency threshold, the event categories within the event cluster are deemed too dispersed and need to be split. The preset minimum category consistency threshold is greater than 0 and less than 1 and is configured by the system administrator based on the diversity of event types in the park. When an event cluster fails any of the above constraints, the system splits the event cluster, removes the events that fail the constraints from the current event cluster, re-performs pairwise correlation calculation and transitive closure merging on the removed events, and forms a new event cluster. The newly formed event cluster also needs to pass the intra-cluster consistency check.

[0052] The reason for setting up intra-cluster consistency checks is that while the transitive closure method can effectively discover indirectly related events, it also carries the risk of over-merging due to the chain bridging effect. The chain bridging effect means that if event A is related to event B, event B is related to event C, and event C is related to event D, then through the transitive closure method, event A and event D will be grouped into the same event cluster. However, event A and event D may be spatially far apart and completely different in event category, actually belonging to different risk sources. The maximum spatial span constraint within the cluster prevents the over-merging of spatially unrelated events by limiting the number of fire zones covered by the event cluster. The intra-cluster centrality constraint identifies marginal events with insufficient relevance to the core of the event cluster by checking the relevance between each event and the central event of the event cluster. The minimum event category consistency constraint within the cluster prevents the mixing and merging of semantically unrelated events by checking the concentration of event categories within the event cluster. These three constraints check the quality of the event cluster from the perspectives of space, correlation strength, and semantic consistency, respectively, to jointly ensure that the construction result of the event cluster neither misses related events nor over-merges irrelevant events.

[0053] After the above merging and verification process, all events in the unified processing queue are divided into several event clusters, each of which includes at least one unified event vector.

[0054] Step 200: Extract multidimensional features for each event cluster, calculate the evidence consistency score based on the multidimensional features, and calculate the true risk score by combining the preset regional basic risk coefficient, as detailed below. Figure 2 As shown.

[0055] Based on the several event clusters output in step 100, calculate the evidence consistency score and the true risk score for each event cluster.

[0056] The evidence consistency score is used to characterize the degree of mutual support and credibility of multi-source evidence within an event cluster for the same real risk. Specifically, the evidence consistency score comprehensively reflects whether multiple wireless nodes within the event cluster point to the same risk type, whether the video review results support the current risk judgment, whether the historical false alarm level of relevant wireless nodes is low, and whether the node communication health status is reliable. The higher the evidence consistency score, the more consistent the evidence from different sources within the event cluster, and the higher the probability that the event cluster is caused by a real fire, a real anomaly, or a real fire risk. The lower the evidence consistency score, the weaker the consistency between the evidence within the event cluster, and the more likely the event cluster is caused by a single point of false alarm, node anomaly, link fluctuation, environmental interference, or human error.

[0057] The true risk score is used to characterize the overall risk level and urgency of handling the risk events corresponding to the event cluster. Specifically, based on the evidence consistency score, the true risk score further combines the regional basic risk coefficient of the area where the event cluster is located, the impact coefficient of key targets, and the event intensity aggregation value to comprehensively reflect whether the current event cluster has a real risk of occurrence, the possible scope and degree of impact and loss after occurrence, and whether it is necessary to prioritize the dispatch of fire-fighting resources for handling. The higher the true risk score, the higher the real risk, the greater the impact, and the more urgent the handling of the event cluster. The lower the true risk score, the lower the risk or the insufficient authenticity of the event cluster, and the more appropriate the handling can be, with priority given to review, observation, or low-level notification.

[0058] The purpose of this step is to quantitatively assess the risk credibility of the event clusters constructed in step 100. In a park fire protection scenario, not all event clusters that trigger alarms correspond to real fire risks. Some event clusters may be caused by non-fire factors such as wireless signal jitter, node hardware failure, environmental interference, or human error. If the system triggers the same level of linkage action for all event clusters indiscriminately, a large amount of linkage resources will be consumed by false alarms, and truly high-risk events may not receive timely responses due to insufficient resources. By introducing two levels of quantitative indicators, namely evidence consistency score and true risk score, the system can comprehensively judge the authenticity of event clusters from multiple independent evidence dimensions, distinguishing high-credibility real risk event clusters from low-credibility suspected false alarm event clusters, and providing a reliable decision-making basis for the generation and prioritization of differentiated linkage actions in subsequent steps.

[0059] For each event cluster, multidimensional features are extracted, including multi-node co-occurrence consistency, video verification support, historical false alarm rate, and node health.

[0060] The reason for selecting the above four dimensions as input features for the evidence consistency score is that these four dimensions provide complementary judgment information on the authenticity of event clusters from different evidence sources and evaluation perspectives. Multi-node co-occurrence consistency assesses whether the responses of multiple independent wireless nodes to the same risk source are consistent at the sensor level, effectively identifying the difference between false alarms from a single wireless node and coordinated responses from multiple wireless nodes. Video verification support provides a second information source independent of sensors at the visual verification level, enabling cross-verification of sensor alarms. Historical false alarm rate reflects the long-term reliability performance of each wireless node in the event cluster from a historical statistical perspective, reducing the impact of frequently false alarming wireless nodes on the overall score. Node health assesses the reliability of event reporting at the network communication level, identifying abnormal events caused by network quality degradation. The information sources of the four dimensions are independent of each other; the failure or absence of any single dimension will not lead to a completely inaccurate overall score, thus ensuring the robustness of the evidence consistency score.

[0061] The multi-node co-occurrence consistency represents the degree of consistency in the responses of different wireless nodes within an event cluster to the same risk source. It is calculated by dividing the number of wireless nodes in the event cluster that point to the same risk type by the total number of wireless nodes in the event cluster. When all wireless nodes in the event cluster point to the same risk type, the multi-node co-occurrence consistency is equal to 1.

[0062] The video verification support level represents the degree to which the video access unit's verification results for the area where the event cluster is located support the actual risk. The verification result information output by the video access unit includes at least one of the following: smoke, suspected fire, no obstruction, missing image, or indeterminate. The value rules are as follows: if the video verification result is smoke or suspected fire, the video verification support level takes the highest value; if the video verification result is no obstruction, the video verification support level takes the middle value; if the video verification result is missing image or indeterminate, the video verification support level takes the lowest value; if no video verification result has been obtained, the video verification support level takes the lowest value. The value range of the video verification support level is between 0 and 1, and the specific value is configured by the system administrator according to the degree of support for the actual risk based on various verification conclusions.

[0063] The historical false alarm rate is the average false alarm rate of each wireless node in the event cluster within a preset historical period. It is calculated by summing the false alarm rates of each wireless node in the event cluster within the preset historical period, and then dividing by the total number of wireless nodes in the event cluster. The false alarm rate of each wireless node refers to the ratio of the number of events ultimately confirmed as false alarms by that node within a rolling time window corresponding to the preset historical period to the total number of events reported by that wireless node within the same rolling time window. The rolling time window is a continuous time interval ending at the current time and spanning the preset historical period; the system only counts events within this time interval. Historical event records outside the specified time interval are not included in the false alarm rate calculation. The system maintains a time-sorted event record queue for each wireless node. Whenever a new closed-loop conclusion is generated, the event record is appended to the end of the queue. At the same time, it checks whether the event record at the head of the queue has exceeded the rolling time window corresponding to the preset historical period. If it has, it is removed from the queue. Through the above rolling time window mechanism, the calculation of the historical false alarm rate is always based on the valid samples within the preset historical period. Expired samples are automatically eliminated, thereby ensuring that the historical false alarm rate can reflect the recent actual performance of the wireless node rather than the cumulative average of all historical data.

[0064] The reason for using a rolling time window-based historical false alarm rate calculation method instead of a full historical accumulation method is that the working status and environmental conditions of wireless nodes in the campus change over time. A wireless node may experience frequent false alarms in the early stages of deployment due to improper installation location or environmental interference, but the false alarm rate will significantly decrease after adjustments. Conversely, a long-term stable wireless node may also experience an increasing false alarm rate recently due to hardware aging or environmental changes. If the full historical accumulation method is used to calculate the false alarm rate, a large number of early false alarm records will inflate the false alarm rate of the wireless node in the long term, and even if the wireless node performs well recently, it will not be reflected in the false alarm rate value, causing the system's reliability assessment of the wireless node to lag behind the actual situation. The rolling time window mechanism automatically eliminates expired samples that exceed the preset historical period, ensuring that the historical false alarm rate always reflects the actual false alarm level of the wireless node in the recent period. It can capture the changing trend of wireless node reliability in a timely manner and provide more timely input data for evidence consistency scoring.

[0065] The preset historical period is configured by the system administrator based on the amount of historical data accumulated in the park and the frequency of events. In parks with a high event frequency, it can be appropriately shortened to improve the timeliness of data, while in parks with a low event frequency, it can be appropriately extended to ensure sufficient sample size.

[0066] The node health score represents the network communication quality status of each wireless node in the event cluster. It is calculated by multiplying the retransmission health component and signal health component of each wireless node in the event cluster to obtain the single-node health score. The retransmission health component is calculated by subtracting the ratio of the number of retransmissions when the wireless node reported the event to a preset maximum number of retransmissions from 1. If the number of retransmissions is greater than the preset maximum number of retransmissions, the retransmission health component is set to 0. The signal health component is calculated by subtracting a preset minimum acceptable signal strength value from the signal strength when the wireless node reported the event, and then dividing by the difference between a preset reference signal strength value and the preset minimum acceptable signal strength value. If the original result is less than 0, the signal health component is set to 0; if the original result is greater than 1, the signal health component is set to 1. Through this method, the signal strength is mapped from the dBm domain to a positive linear interval between 0 and 1. The closer the signal strength is to the preset reference signal strength value, the closer the signal health component is to 1. The closer the signal health component is to or below the preset minimum acceptable signal strength value, the closer it is to 0. The health value of each node in the event cluster is obtained by summing the individual node health values ​​of all wireless nodes in the event cluster and then dividing by the total number of wireless nodes in the event cluster. The preset maximum retransmission count is configured by the system administrator according to the retransmission mechanism of the wireless networking protocol; the value can be differentiated according to the protocol characteristics under different protocol environments. The preset reference signal strength value represents the signal strength benchmark under ideal communication conditions and is configured by the system administrator according to the actual deployment environment of the campus wireless network and the transmission power of the wireless nodes. The preset minimum acceptable signal strength value represents the lower limit of signal strength at which the wireless node's communication quality drops to an unusable level and is configured by the system administrator according to the actual receiving sensitivity of the campus wireless network; the preset minimum acceptable signal strength value must be less than the preset reference signal strength value. Through the above calculation method, the value of a single node health value is always between 0 and 1. When the wireless node's communication quality is good, the node health value is close to 1; when the wireless node retransmits frequently or the signal is weak, the node health value is close to 0.

[0067] The reason for calculating the health value of a single node by multiplying the retransmission health component and the signal health component instead of adding them is that the number of retransmissions and signal strength are two key but independent indicators reflecting the communication quality of a wireless node. A serious deterioration in either one should lead to a significant decrease in the health value of the wireless node. If the addition method is used, when one indicator is seriously deteriorated while the other is performing well, the weighted sum of the two may still produce a moderately high health value, masking the fact that the node has serious shortcomings in communication quality. By multiplying, when either component approaches 0, the health value of the single wireless node also approaches 0, thus ensuring that a wireless node with serious defects in any dimension of communication quality will be assigned a lower health value. This appropriately reduces the contribution of the events reported by the wireless node to the evidence consistency score, avoiding the misleading influence of unreliable wireless nodes on the risk assessment results.

[0068] Based on the aforementioned multidimensional features, an evidence consistency score for an event cluster is calculated. This score is equal to the sum of four factors: the product of the co-occurrence consistency of multiple nodes and a preset first weighting coefficient, the product of the video review support and a preset second weighting coefficient, the product of the difference between 1 and the historical false alarm rate and a preset third weighting coefficient, and the product of the node health and a preset fourth weighting coefficient. The sum of the four weighting coefficients (preset first, second, third, and fourth) equals 1, and each coefficient is greater than 0. The values ​​of these four weighting coefficients are all greater than 0 and less than 1. The specific values ​​are configured by the system administrator based on the park's video coverage and wireless node deployment density. In parks with high video coverage, the preset second weighting coefficient can be appropriately increased; in parks with high wireless node deployment density, the preset first weighting coefficient can be appropriately increased. During adjustments, the sum of the four weighting coefficients must always equal 1. The evidence consistency score ranges from 0 to 1; a higher value indicates greater consistency of evidence within the event cluster, and a higher credibility of the actual risk corresponding to that event cluster.

[0069] The advantage of using a four-dimensional weighted summation method to calculate the evidence consistency score lies in its ability to integrate four independent evidence dimensions from the sensor layer, video layer, historical statistics layer, and network communication layer into a comprehensive credibility index. This allows the system to intuitively quantify and compare the evidence consistency of event clusters on a single numerical basis. The weighted summation method also offers the advantage of flexible adjustment of the contribution of each dimension through weight coefficients. When a particular source of evidence is more reliable or important in a specific area, the system administrator can increase the corresponding weight coefficient to enhance the influence of that dimension on the overall score without modifying the scoring algorithm itself. Furthermore, since the sum of the four weight coefficients is always equal to 1 and each weight coefficient is greater than 0, the evidence consistency score is strictly constrained to a range of 0 to 1. This facilitates comparison and judgment with various thresholds in subsequent steps and also makes it convenient to use as input for further weighted calculations in the calculation of the true risk score.

[0070] Based on the evidence consistency score, the true risk score of the event cluster is calculated by further combining the regional basic risk coefficient, the key target influence coefficient, and the event intensity aggregation value. The true risk score is equal to the sum of the following four products: the evidence consistency score multiplied by the preset sixth weight coefficient, the regional basic risk coefficient multiplied by the preset seventh weight coefficient, the key target influence coefficient multiplied by the preset eighth weight coefficient, and the event intensity aggregation value multiplied by the preset ninth weight coefficient. The sum of the four preset weight coefficients (sixth, seventh, eighth, and ninth) is equal to 1, and each weight coefficient is greater than 0. The values ​​of the above four weight coefficients are all greater than 0 and less than 1. Their specific values ​​are configured by the system administrator according to the park's risk management strategy. In scenarios that place greater emphasis on evidence reliability, the preset sixth weight coefficient can be appropriately increased. In parks with a large number of high-value equipment or densely populated areas, the preset eighth weight coefficient can be appropriately increased. When adjusting, it is necessary to ensure that the sum of the four weight coefficients is always equal to 1.

[0071] The basic risk coefficient of the area is a quantitative value of the inherent risk level of the fire compartment where the event cluster is located. It is pre-configured in the park spatial mapping information and ranges from 0 to 1. The larger the value, the higher the inherent risk level of the area. The initial value of the coefficient is configured by the system administrator according to the functional attributes of each fire compartment, the fire hazard level of the stored items, and historical fire records. The initial value of the storage area for flammable and explosive materials should be set to a higher level, and the initial value of the ordinary office area should be set to a lower level. The coefficient is adaptively updated in step 500 based on the closed-loop result.

[0072] The key object impact coefficient is a quantitative value of the density of key equipment or key personnel in the area where the event cluster is located. The value ranges from 0 to 1. The larger the value, the more key objects in the area are affected or the more sensitive they are. The coefficient is configured by the system administrator according to the value level of key production equipment, personnel density and evacuation difficulty in each area. Areas containing core equipment or hazardous materials storage facilities should be set to a higher level, and ordinary passage areas should be set to a lower level.

[0073] The event intensity aggregation value is the aggregation result of the event intensity of all events in the event cluster; it is calculated by taking the maximum value of the event intensity of all events in the event cluster as the event intensity aggregation value of the event cluster.

[0074] In the aforementioned true risk score, the regional basic risk coefficient, the key object impact coefficient, and the event intensity aggregate value are used to characterize the regional background risk of the event cluster, the sensitivity of the affected objects, and the severity of the current event, respectively. Among them, the regional basic risk coefficient represents the inherent risk benchmark formed by the fire protection zone where the event cluster is located due to its functional use, storage of combustibles or hazardous materials, historical fire records, personnel activity characteristics, and fire protection facility conditions, without considering real-time evidence of the current event. The key object impact coefficient represents the importance and sensitivity of the key production equipment, important materials, densely populated areas, or objects with difficult evacuation that may be affected by the event cluster. The event intensity aggregate value represents the upper limit of the real-time abnormal intensity reflected by all alarms, faults, manual triggers, or other related events within the event cluster, and is used to highlight the contribution of the most severe event in the event cluster to the true risk score.

[0075] The reason for using the maximum value instead of the average or summation value as the event intensity aggregation method is that the risk urgency of an event cluster in a fire scenario should be determined by the most severe single event, not by the average level of all events. If the average value method is used, when an event cluster contains one alarm event with extremely high event intensity and multiple auxiliary events with low event intensity, the average value will be lowered by a large number of low-intensity events, causing the system to underestimate the actual risk urgency of the event cluster. If the summation value method is used, the more events contained in the event cluster, the larger the aggregation value, which may lead to an event cluster containing a large number of low-intensity events obtaining a higher aggregation value than an event cluster containing a small number of high-intensity events, which is inconsistent with the actual risk urgency. The maximum value method ensures that the event intensity aggregation value always reflects the intensity level of the most severe event in the event cluster, so that an event cluster containing at least one high-intensity alarm event can obtain an aggregation value that matches its actual risk urgency.

[0076] The true risk score ranges from 0 to 1. The higher the value, the higher and more urgent the true risk corresponding to the event cluster. Through this step, the system can distinguish between highly reliable fires that are consistently pointed to by multiple sources and pseudo-high-priority events caused by wireless jitter, node anomalies, or single-point false alarms.

[0077] Step 300: Generate a set of candidate linked actions for each event cluster based on the evidence consistency score and the true risk score; construct an action execution sequence graph based on the set of candidate linked actions and solve for the priority execution sequence; summarize the priority execution sequences based on the action execution sequence graph to obtain the total execution queue, as detailed below. Figure 2 As shown.

[0078] Based on the several event clusters output in step 200 and their corresponding evidence consistency scores and true risk scores, a set of candidate linkage actions is generated for each event cluster, and an action execution sequence diagram is constructed to solve the priority execution sequence.

[0079] The purpose of this step is to transform the quantified risk assessment results in step 200 into specific action execution plans. In the scenario of fire linkage in the park, event clusters with different risk levels and evidence consistency levels need to trigger linkage actions of different scopes and intensities. At the same time, there are resource competition and execution constraints between the linkage actions of multiple event clusters, and the optimal execution order needs to be determined through global optimization. This step models the dependencies, mutual exclusions, and parallelizability between candidate linkage actions as a directed acyclic graph, and combines the action priority value to solve the layer-by-layer topology sorting based on a greedy strategy. Under the premise of satisfying all constraints, it can quickly obtain the priority execution sequence that maximizes the global position weighted benefit, ensuring that high-priority linkage actions can be executed in the shortest time, while avoiding violations of business logic constraints.

[0080] The set of candidate linkage actions consists of several candidate linkage actions corresponding to the event cluster. The candidate linkage actions include notification, review, and escalation actions. Notification actions include notification to the responsible person, dispatching of orders by on-duty personnel, and on-site confirmation instructions. Review actions include video review requests and manual review holding. Escalation actions include local audio-visual linkage, regional broadcast prompts, and escalation reports from higher authorities.

[0081] The rules for generating candidate action sets are related to the evidence consistency score and the true risk score of the event cluster, and the specific rules are as follows:

[0082] When the true risk score and evidence consistency score of an event cluster are both greater than or equal to their respective preset high thresholds, the event cluster is determined to be a high-risk, high-consistency event cluster. Specifically, when the true risk score of an event cluster is greater than or equal to the preset true risk score high threshold and the evidence consistency score is greater than or equal to the preset evidence consistency score high threshold, a set of candidate linkage actions is generated, including review, notification, and escalation actions. These actions include video review requests, notifications to responsible persons, dispatching of orders by on-duty personnel, on-site confirmation instructions, local audio-visual linkage, regional broadcast prompts, and escalation reports from higher authorities.

[0083] When the true risk score of an event cluster is greater than or equal to the preset high threshold for true risk score, but the evidence consistency score is less than the preset high threshold for evidence consistency score, or the true risk score is between the preset high threshold for true risk score and the preset low threshold for true risk score, the event cluster is determined to be a medium-risk event cluster. Specifically, when the true risk score of an event cluster is greater than or equal to the preset high threshold for true risk score but the evidence consistency score is less than the preset high threshold for evidence consistency score, or when the true risk score of an event cluster is greater than or equal to the preset low threshold for true risk score and less than the preset high threshold for true risk score, a set of candidate linkage actions, including review and notification actions, is generated. These actions include video review requests, notifications to responsible persons, dispatching of duty personnel, and on-site confirmation instructions.

[0084] When the true risk score of an event cluster is lower than the preset low threshold for true risk score, the event cluster is determined to be a low-confidence event cluster. Specifically, when the true risk score of an event cluster is lower than the preset low threshold for true risk score, a set of candidate linkage actions including review actions is generated, specifically including video review requests and manual review holding.

[0085] The preset high threshold for true risk score is greater than the preset low threshold for true risk score and less than or equal to 1. The preset low threshold for true risk score is greater than 0 and less than the preset high threshold for true risk score. The preset high threshold for evidence consistency score is greater than 0 and less than or equal to 1. The above three thresholds are configured by the system administrator according to the park's fire safety management strategy and the acceptable false alarm tolerance. In parks with extremely high safety requirements, the preset high threshold for true risk score and the preset low threshold for true risk score can be appropriately reduced to expand the triggering range of high-level responses. In parks with frequent false alarms, the above thresholds can be appropriately increased to reduce unnecessary candidate linkage actions.

[0086] Based on the above rules, event clusters with different risk levels and levels of evidence consistency correspond to different ranges of candidate linkage action sets. The candidate linkage action set corresponding to high-risk, high-consistency event clusters includes higher-level candidate linkage actions, while low-credibility event clusters mainly consist of review and observation-type candidate linkage actions.

[0087] The reason for adopting a tiered generation rule based on real risk scores and evidence consistency scores is that the execution of fire-fighting linkage actions is irreversible and resource-intensive. Once a high-level linkage action is triggered, it will have a significant social impact and resource consumption. If a candidate set containing all types of linkage actions is generated indiscriminately for all event clusters, even low-credibility event clusters may trigger high-level actions such as local sound and light linkage or escalation reporting to higher authorities, causing unnecessary social panic and resource waste. By dividing event clusters into three levels—high-risk and high-consistency, medium-risk, and low-credibility—according to real risk scores and evidence consistency scores, and configuring a different range of candidate linkage action sets for each level, the system can ensure that the intensity of linkage actions matches the risk level and evidence credibility of the event cluster. Low-credibility event clusters only generate review actions, allowing the system to prioritize collecting more evidence rather than rashly taking strong measures in uncertain situations. High-risk and high-consistency event clusters generate a full range of candidate sets, including escalation actions, ensuring that real high-risk events can obtain the most comprehensive linkage response.

[0088] After generating the candidate linkage action set, an action execution sequence graph is constructed for each event cluster by combining linkage resource status information and candidate linkage action constraint information. The action execution sequence graph is a directed acyclic graph, consisting of a candidate linkage action node set and a directed edge set. The candidate linkage action node set is the candidate linkage action set for that event cluster, and the directed edge set represents the sequential dependencies between candidate linkage actions. If a candidate linkage action must be executed before another candidate linkage action, a directed edge is established in the directed edge set from the former to the latter, indicating that the execution of the latter is premised on the completion of the former. The linkage resource status information includes currently available on-duty personnel, available terminals, triggerable control modules, occupied linkage channels, and the set of candidate linkage actions currently being executed. The candidate linkage action constraint information includes at least the sequential dependencies, mutual exclusion relationships, parallelism relationships, and forced preemption conditions between candidate linkage actions.

[0089] The establishment of the directed edges is based on the pre- and post-dependent relationships in the candidate linkage action constraint information. The pre- and post-dependent relationships refer to the fact that some candidate linkage actions must be executed logically before others, based on the business logic of fire linkage. At the same time, the candidate linkage action constraint information also includes mutual exclusion relationships and parallelizable relationships. Mutual exclusion relationships mean that two candidate linkage actions cannot be executed at the same time and must be executed in sequence. Parallelizable relationships mean that two candidate linkage actions have no dependency relationship and are not mutually exclusive, and can be executed at the same time.

[0090] The purpose of the action execution sequence diagram is to structurally represent the dependencies, mutual exclusions, and parallelizability among candidate linked actions in the candidate linked action set in the form of a directed acyclic graph, providing a constraint basis for subsequent solutions to the priority execution sequence. Specifically, the set of directed edges in the action execution sequence diagram defines the inviolable execution order among candidate linked actions. Any priority execution sequence must satisfy the predecessor-successor relationship specified by the set of directed edges, that is, the execution batch number of the predecessor candidate linked action in the priority execution sequence must be less than or equal to the execution batch number of the successor candidate linked action. The implicit mutual exclusion relationship in the action execution sequence diagram further constrains candidate linked action pairs that cannot be arranged simultaneously within the same execution batch, while the parallelizability relationship indicates... The system identifies candidate linked actions that can be executed simultaneously within the same execution batch. Through the action execution sequence diagram, the system can identify currently executable candidate linked actions layer by layer according to the topological structure of the directed acyclic graph during the subsequent layer-by-layer topological sorting process. In each layer, it performs a greedy selection based on action priority values ​​and mutual exclusion relationships, thereby obtaining the priority execution sequence that maximizes the global position-weighted benefit while satisfying all constraints. In addition, the dependencies defined by the action execution sequence diagram also play a constraining role in the subsequent park-level aggregation and overall execution queue construction process, ensuring that when the priority execution sequences of multiple event clusters are merged into the overall execution queue, the execution order of candidate linked actions within each event cluster as specified by the action execution sequence diagram is not disrupted.

[0091] Calculate an action priority value for each candidate linkage action; the action priority value is the sum of the product of the actual risk score of the event cluster and the preset first priority value weight coefficient, the product of the timeliness benefit of the candidate linkage action and the preset second priority value weight coefficient, and the product of the preset evidence clarification gain of the candidate linkage action and the preset third priority value weight coefficient, and then subtract the product of the resource occupation cost of the candidate linkage action and the preset fourth priority value weight coefficient.

[0092] The reason for using four factors—real risk score, timeliness benefit, evidence clarification gain, and resource consumption cost—to calculate the action priority is that the execution priority of linked actions should not be determined solely by the risk level of the event cluster. Instead, it should comprehensively consider the urgency of the action at the current moment, its informational contribution to subsequent decision-making, and its consumption of limited linked resources. The real risk score reflects the urgency of the event cluster itself, ensuring that linked actions for high-risk event clusters receive a higher basic priority. Timeliness benefit reflects the time sensitivity of linked actions, ensuring that time-sensitive actions are prioritized before the time window closes. Evidence clarification gain reflects the informational value of linked actions to subsequent decision-making, ensuring that actions that can provide key evidence to the system are prioritized, thereby accelerating the risk confirmation process of the event cluster. Resource consumption cost is introduced as a deduction, so that when action priority values ​​are similar, actions with lower resource consumption are prioritized, thereby saving linked resources while ensuring response effectiveness and reserving resource margins for potentially higher-risk events later.

[0093] The preset first priority value weight coefficient, preset second priority value weight coefficient, preset third priority value weight coefficient, and preset fourth priority value weight coefficient are all greater than 0. The values ​​of the above four weight coefficients are all positive real numbers greater than 0. Their specific values ​​are configured by the system administrator according to the emphasis of the park linkage strategy. In scenarios that focus more on rapid response, the preset second priority value weight coefficient can be appropriately increased. In scenarios where linkage resources are scarce, the preset fourth priority value weight coefficient can be appropriately increased to strengthen the resource conservation orientation.

[0094] The timeliness benefit refers to the additional benefit that can be obtained by executing the candidate linkage action at the current moment compared to delaying its execution. The value ranges from 0 to 1, with a higher value indicating greater time urgency. The specific value of the timeliness benefit is calculated by the system based on the preset timeliness benchmark values ​​for various candidate linkage actions and the duration of the current event cluster. The preset timeliness benchmark values ​​are pre-configured by the system administrator for each type of candidate linkage action. Action types with rapid suppression value should be configured with higher preset timeliness benchmark values, while action types focused on observation and maintenance should be configured with lower preset timeliness benchmark values. As the duration of the event cluster increases, the timeliness benefit decays according to an exponential function with the preset timeliness benchmark value multiplied by the natural constant as the base and the event cluster duration divided by the negative value of the preset timeliness decay period as the exponent. The preset timeliness decay period is configured by the system administrator based on the timeliness requirements of the park's fire response. The gain for subsequent evidence clarification indicates how much new evidence the candidate linkage action can provide to the system to assist subsequent decision-making after execution, with a value range from 0 to 1. The specific value of the evidence clarification gain is calculated by the system based on the preset timeliness benchmark value multiplied by the natural constant as the base and the event cluster duration divided by the negative value of the preset timeliness decay period as the exponent. The preset evidence clarification gain benchmark value is pre-configured by the system administrator for each candidate linkage action type. Action types that can directly obtain on-site information should be configured with a higher preset evidence clarification gain benchmark value, while action types that are mainly for execution control should be configured with a lower preset evidence clarification gain benchmark value. The resource occupation cost represents the amount of linkage resources required to execute the candidate linkage action, with a value ranging from 0 to 1. The larger the value, the more resources the candidate linkage action occupies. The specific value of the resource occupation cost is calculated by the system based on the preset resource occupation cost benchmark value and the current availability ratio of linkage resources for each type of candidate linkage action. The preset resource occupation cost benchmark value is pre-configured by the system administrator for each candidate linkage action type. Action types that require manpower should be configured with a higher preset resource occupation cost benchmark value, while action types that only require system channels should be configured with a lower preset resource occupation cost benchmark value. When the availability ratio of linkage resources decreases, the resource occupation cost increases accordingly. The preset resource occupation cost benchmark value is adaptively adjusted in step 500 based on the actual resource consumption.

[0095] After constructing the action execution sequence diagram and calculating the action priority value of each candidate linked action, the priority execution sequence is solved for all candidate linked actions of the event cluster, under the premise of satisfying the dependencies and mutual exclusion relationships. Since the time efficiency benefit of each candidate linked action decays with the actual execution time, the earlier a candidate linked action is placed in the priority execution sequence, the higher its time efficiency benefit. Therefore, the actual benefit of a candidate linked action is directly related to its position in the priority execution sequence. The objective of solving the priority execution sequence is to maximize the sum of the position-weighted benefits of all candidate linked actions in the priority execution sequence, under the premise of satisfying all constraints. The position-weighted benefit is the product of the action priority value of each candidate linked action and the position decay factor of that candidate linked action in the priority execution sequence. The position decay factor is calculated by using the natural constant as the base and the position decay factor of the candidate linked action in the priority execution sequence. The exponent is the negative value of the quotient obtained by multiplying the execution batch number of an action in the priority execution sequence by a preset batch interval duration and then dividing by a preset time-effect decay period. The preset batch interval duration is the estimated time interval between two adjacent execution batches, configured by the system administrator based on the average execution time of candidate linked actions. By introducing a position decay factor, candidate linked actions with higher priority and stronger timeliness can obtain greater position-weighted benefits when placed earlier, thus directly linking the solution objective to the execution order. The constraints are: for each directed edge in the directed edge set, the execution batch number of the preceding candidate linked action in the priority execution sequence is less than or equal to the execution batch number of the succeeding candidate linked action; for each pair of mutually exclusive candidate linked actions in the candidate linked action constraint information, they are not arranged in the same parallel execution batch in the priority execution sequence.

[0096] The solution process employs a greedy, layer-by-layer topological sorting method. Specifically, firstly, the in-degree of each candidate linked action node in the directed acyclic graph is calculated, i.e., the number of directed edges pointing to that node. All candidate linked action nodes with an in-degree of 0 are added to the current selectable set as the candidate range for the first execution batch. Then, in the current selectable set, candidate linked actions are selected in descending order of action priority to be added to the current execution batch. During selection, mutual exclusion relationships must be checked. If the candidate linked action to be selected has a mutual exclusion relationship with any existing candidate linked action in the current execution batch, then the candidate linked action is not added to the current execution batch but is retained in the selectable set to wait for subsequent batches. All candidate linked actions in the current selectable set are then added to the current batch. After all the selected actions have undergone the above selection and judgment, the current execution batch is determined. Then, all candidate actions in the current execution batch are removed from the directed acyclic graph, and the in-degree of the remaining nodes is updated. Newly generated nodes with an in-degree of 0 are added to the next round of the selectable set. The above process is repeated until all candidate actions are assigned to a certain execution batch. Within the same execution batch, candidate actions that do not have mutual exclusion relationships are marked as parallel execution. Through the above layer-by-layer greedy selection method, the system can obtain the priority execution sequence that satisfies all constraints in a single traversal, without enumerating all legal topological sorting results. The computational complexity is linearly related to the number of candidate actions and the number of directed edges, which is suitable for real-time linkage control scenarios.

[0097] The reason for adopting a greedy strategy-based layer-by-layer topology sorting method instead of exhaustive search or dynamic programming to solve the priority execution sequence is that the fire-fighting linkage scenario in the park has strict real-time requirements for solution speed. The system needs to complete the entire processing flow from event reception to linkage action issuance within seconds. Although the exhaustive search method can find the global optimum, its computational complexity increases exponentially with the number of candidate linkage actions, which cannot meet the real-time requirements when the number of candidate linkage actions is large. Although the dynamic programming method can solve in polynomial time, it requires the construction and maintenance of a complex state space, which is difficult to implement and consumes a lot of memory. The greedy strategy-based layer-by-layer topology sorting method selects the candidate linkage action with the highest priority value in the current set of available actions at each layer and adds it to the execution batch. Its computational complexity is linearly related to the number of candidate linkage actions and the number of directed edges, and it can complete the solution in a very short time. Although the greedy strategy does not guarantee a global optimum, since the dependencies between candidate linkage actions in the fire-fighting linkage scenario are usually hierarchical, the greedy strategy can obtain near-global optimum solution results in most practical scenarios, achieving a good balance between solution quality and solution speed.

[0098] The priority execution sequences of all event clusters are aggregated at the park level to obtain the overall execution queue. The park-level aggregation rule is as follows: unexecuted candidate linkage actions in all event clusters are arranged in descending order of their action priority value, while maintaining the sequential dependencies defined by the set of directed edges in the action execution sequence diagram within each event cluster. That is, for any pair of predecessor and successor candidate linkage actions connected by directed edges in the action execution sequence diagram, the predecessor candidate linkage action must be arranged before the successor candidate linkage action in the overall execution queue. When candidate linkage actions from different event clusters need to occupy the same linkage resource, they are preferentially allocated to candidate linkage actions with higher action priority values. Through the above park-level aggregation rule, all constraints encoded in the action execution sequence diagram are fully preserved in the overall execution queue. This ensures that the overall execution queue reflects both the global action priority value sorting across event clusters and strictly adheres to the execution order and mutual exclusion constraints specified by the action execution sequence diagram within each event cluster, ensuring that the actual execution of linkage actions does not violate the business logic of fire linkage.

[0099] Step 400: Output and issue the first batch of candidate linkage actions according to the total execution queue. Within a preset listening time, update the evidence consistency score and the true risk score based on the feedback information, and partially rearrange the remaining candidate linkage actions to obtain the updated total execution queue. Execute subsequent candidate linkage actions according to the updated total execution queue, as follows: Figure 3 As shown.

[0100] Based on the total execution queue output in step 300, the first batch of candidate linkage actions are output in the order of the total execution queue; the first batch of candidate linkage actions refers to the set of candidate linkage actions that are at the top of the total execution queue and can be satisfied by the current linkage resources.

[0101] The purpose of this step is to realize the actual issuance and execution of linkage actions and the dynamic reordering mechanism based on feedback information. Traditional fire linkage systems usually trigger all linkage actions at once, and the sequence of linkage actions is fixed once determined and cannot be adjusted according to new information obtained during execution. However, in actual fire scenarios, after the first batch of linkage actions are issued, the system will receive feedback information such as video review results, manual confirmation, continuous alarms or recovery of nodes. This feedback information may change the risk assessment results of the event cluster. This step continuously receives feedback information by setting up feedback listening points and converts the feedback information into incremental evidence to update the evidence consistency score and the true risk score. Then, it recalculates the action priority value of the candidate linkage actions that have not yet been executed and performs local reordering, so that the execution order of linkage actions can be continuously optimized as new evidence arrives, realizing adaptive sequence control that can be reordered during execution.

[0102] Based on the risk level and the consistency of evidence of the event clusters, the execution strategies for the first batch of candidate linkage actions are differentiated.

[0103] For high-risk, high-consistency event clusters, i.e., event clusters where the actual risk score is greater than or equal to the preset high threshold for actual risk score and the evidence consistency score is greater than or equal to the preset high threshold for evidence consistency score, priority is given to executing candidate linkage actions in the total execution queue that belong to the event cluster and have rapid suppression value and high confidence notification value, including local sound and light linkage, regional broadcast prompts, notification of responsible persons, and escalation reporting to higher authorities.

[0104] For medium-risk event clusters, i.e., event clusters whose actual risk score is greater than or equal to the preset low threshold for actual risk score and less than the preset high threshold for actual risk score, or event clusters whose actual risk score is greater than or equal to the preset high threshold for actual risk score but whose evidence consistency score is less than the preset high threshold for evidence consistency score, priority will be given to executing candidate linkage actions such as video review requests, notifications to responsible persons, and on-site confirmation instructions belonging to that event cluster in the overall execution queue.

[0105] For low-reliability event clusters, i.e. event clusters whose true risk score is less than the preset low threshold of true risk score, video review requests and manual review keep-candidate linkage actions belonging to the event cluster in the total execution queue are executed first, so that the event cluster enters the observation and review channel.

[0106] After the first batch of candidate linkage actions are issued, a feedback listening point is set up for each event cluster. The feedback listening point refers to the information receiving interface established by the system for each event cluster, which is used to continuously listen for feedback information within a preset listening time. The feedback information includes video review results, manual confirmation, continuous alarm information from nodes, node recovery reporting information, on-site confirmation results, and status reports on whether the candidate linkage action was successfully executed. The preset listening time is a positive real number greater than 0. The preset listening time is configured by the system administrator based on the average feedback cycle of the park's fire response process. It can be appropriately shortened in parks with faster manual response speeds and appropriately extended in scenarios requiring on-site confirmation and involving long distances. If no feedback information is received within the preset listening time, the system continues to execute subsequent candidate linkage actions in the order of the overall execution queue.

[0107] If a feedback monitoring point receives new feedback information within a preset monitoring period, the new feedback information is mapped as incremental evidence. The new feedback information refers to feedback events related to the event cluster that the feedback monitoring point receives through the information receiving interface after the first batch of candidate linkage actions are issued and before the preset monitoring period expires. The sources of the new feedback information include video access units, personnel message push terminals, on-site handling terminals, linkage control hosts, and wireless nodes in the park's wireless network. The new feedback information includes the following types according to its source and content.

[0108] The video review result feedback information refers to the review conclusion returned by the video access unit after receiving the video review request and analyzing and processing the video image of the area where the event cluster is located. The value range of the review conclusion is consistent with the review result information on which the video review support is based in step 200, including at least one of the following: smoke, suspected fire, no obstruction, missing image, or indeterminate. The video review result feedback information includes the device identifier of the video access unit, the review conclusion, the review timestamp, and the identifier of the video segment on which the review is based.

[0109] Manually generated feedback information refers to confirmation information returned by the responsible person, on-duty personnel, or on-site handling personnel after responding to a issued candidate linkage action through a personnel message push terminal or on-site handling terminal. Manually generated feedback information includes the identity identifier of the person generating the feedback, the feedback timestamp, the feedback content, and the candidate linkage action identifier corresponding to the feedback. The feedback content includes at least one of the following: notification received, arrival at the scene, confirmed as a real risk, confirmed as a false alarm, need for reinforcement, and unable to arrive.

[0110] The continuous alarm feedback information refers to alarm events reported by wireless nodes in the event cluster after the first batch of candidate linkage actions have been issued. This type of feedback information indicates that the risk source that triggered the event cluster is still continuously generating abnormal signals that can be detected by sensors. Its event category is alarm, and the event intensity is obtained by the node recalculating the normalization based on the deviation of the latest sensor reading from the preset alarm threshold. The continuous alarm feedback information includes the node identifier that reported the event, the area identifier, the event timestamp, the event intensity, and a set of network quality parameters. Its format is consistent with the unified event vector set in step 100.

[0111] Node recovery reporting feedback information refers to recovery events actively reported by wireless nodes in the event cluster that were originally in an alarm state after the sensor readings have returned to the normal range. This type of feedback information indicates that the environmental parameters monitored by the node have returned from an abnormal state to a normal level. Its event category is recovery, and the event intensity is obtained by the node through normalization calculation based on the proximity of the current sensor reading to the normal range benchmark value. When the sensor reading has fully recovered to the normal range, the event intensity is 0. The node recovery reporting feedback information includes the node identifier that reported the event, the area identifier, the event timestamp, the event intensity, and the set of network quality parameters. Its format is consistent with the unified event vector set in step 100.

[0112] On-site confirmation result feedback information refers to the actual situation reported by on-site personnel through the on-site handling terminal after arriving at the area where the incident cluster is located. On-site confirmation result feedback information includes the on-site personnel's identification, confirmation timestamp, on-site confirmation conclusion, and on-site description information. The on-site confirmation conclusion includes at least one of the following: confirmed fire, confirmed smoke, confirmed equipment malfunction, confirmed false alarm, and normal situation. The on-site description information is a supplementary explanation of the actual situation on-site entered by the on-site personnel in text form.

[0113] The candidate linkage action execution status report feedback information refers to the execution result information returned by the linkage control host after executing the candidate linkage action; this type of feedback information includes the executed candidate linkage action identifier, execution timestamp, execution status, and execution exception description; the execution status includes at least one of execution success, execution failure, and execution timeout; when the execution status is execution failure or execution timeout, the execution exception description records the specific reasons for the execution failure or timeout, including but not limited to target device offline, control channel occupied, instruction issuance timeout, and device response abnormality.

[0114] After being received by the feedback monitoring point, all the aforementioned new feedback information is converted according to the unified event vector format set in step 100. The event timestamps in the converted feedback event vectors are corrected using the same two-level clock correction mechanism as in step 100. The node identifier is the unique code of the device or terminal that generated the feedback information, and the area identifier is the fire zone code corresponding to the device or terminal in the park spatial mapping information. The event category is mapped to the corresponding category among alarm, fault, recovery, node offline, manual trigger, and inspection anomaly according to the type of feedback information. The event intensity is normalized and quantified according to the content of the feedback information. The network quality parameter set records the network routing hop count, response status, retransmission count, and signal quality indicators during the transmission of the feedback information. The original payload summary records the summary of key fields in the original data frame of the feedback information. The incremental evidence is the set composed of all the converted feedback event vectors.

[0115] The reason for uniformly converting all types of feedback information into feedback event vectors with the same format as the unified event vector in step 100 is that the feedback information comes from various heterogeneous devices such as video access units, personnel message push terminals, on-site handling terminals, and linkage control hosts, and their original data formats are different. If the system designs an independent scoring update algorithm for each type of feedback information, the complexity and maintenance cost of the algorithm will increase linearly with the increase of the types of feedback information. By uniformly converting all feedback information into a standardized feedback event vector format, the system can reuse the existing multidimensional feature extraction and scoring calculation logic in step 200 to process incremental evidence. There is no need to develop a separate scoring update algorithm for each type of feedback information, which reduces the complexity of system implementation and ensures the consistency between the scoring update logic and the initial scoring calculation logic.

[0116] Based on the incremental evidence, calculate the incremental consistency score of evidence and the incremental true risk score for the event cluster.

[0117] The calculation method for the incremental evidence consistency score is as follows: the feedback information in the incremental evidence is updated to the multi-dimensional feature quantities of the event cluster, and the updated multi-node co-occurrence consistency, video review support, historical false alarm rate, and node health are recalculated. Then, in the same way as in step 200, the following four items are summed: the product of the updated multi-node co-occurrence consistency with the preset first weight coefficient, the product of the updated video review support with the preset second weight coefficient, the product of the difference between 1 and the updated historical false alarm rate with the preset third weight coefficient, and the product of the updated node health with the preset fourth weight coefficient. The sum of these four items is then subtracted from the original evidence consistency score calculated in step 200. The resulting difference is the incremental evidence consistency score. The preset first to preset fourth weight coefficients are the weight coefficients set in step 200.

[0118] The calculation method for the incremental true risk score is as follows: the original evidence consistency score is added to the incremental evidence consistency score to obtain the updated evidence consistency score. Then, in the same manner as in step 200, the product of the updated evidence consistency score and the preset sixth weight coefficient, the product of the regional basic risk coefficient and the preset seventh weight coefficient, the product of the key object influence coefficient and the preset eighth weight coefficient, and the product of the event intensity aggregation value updated with incremental evidence and the preset ninth weight coefficient are summed. The summation of these four products is then subtracted from the original true risk score calculated in step 200. The difference obtained is the incremental true risk score. The preset sixth to preset ninth weight coefficients are the weight coefficients set in step 200.

[0119] The evidence consistency score increment represents the change in the degree of evidence consistency caused by the new feedback information relative to the original event cluster. When the evidence consistency score increment is positive, it indicates that the new feedback enhances the consistency among multiple sources of evidence and increases the credibility of the event cluster as a real risk. When the evidence consistency score increment is negative, it indicates that the new feedback weakens the consistency among multiple sources of evidence and reduces the credibility of the event cluster as a real risk. The real risk score increment represents the amount of correction caused by the new feedback information to the actual risk level of the event cluster. When the real risk score increment is positive, it indicates that the new feedback increases the risk level, urgency, or handling priority of the event cluster. When the real risk score increment is negative, it indicates that the new feedback decreases the risk level, urgency, or handling priority of the event cluster.

[0120] Based on this, update the evidence consistency score and true risk score of the event cluster. The updated evidence consistency score is equal to the original evidence consistency score plus the evidence consistency score increment; the updated true risk score is equal to the original true risk score plus the true risk score increment.

[0121] Subsequently, based on the updated real risk score, the action priority value of each candidate linkage action that has not yet been executed in the event cluster is recalculated. For a candidate linkage action that has not yet been executed, its updated action priority value is equal to the sum of the product of the updated real risk score and the preset first priority value weight coefficient, the product of the timeliness benefit of the candidate linkage action and the preset second priority value weight coefficient, and the product of the gain of the candidate linkage action on subsequent evidence clarification and the preset third priority value weight coefficient, minus the product of the resource occupation cost of the candidate linkage action and the preset fourth priority value weight coefficient. Among them, the preset first priority value weight coefficient to the preset fourth priority value weight coefficient are the weight coefficients set in step 300.

[0122] While maintaining the dependencies and mutual exclusion relationships, the remaining candidate linkage action sequences of the event cluster are partially rearranged to obtain the updated candidate linkage action sequences. The partial rearrangement refers to rearranging only the candidate linkage actions that have not yet been executed according to the updated action priority value, while the candidate linkage actions that have already been executed are not affected.

[0123] The reason for using local reordering instead of global reordering is that the executed candidate linkage actions have already produced actual effects and their execution results are irreversible. Reordering the executed candidate linkage actions is meaningless and may lead to inconsistencies between the execution records and the actual execution order. Local reordering only adjusts the priority of candidate linkage actions that have not yet been executed, which ensures the integrity of the execution records of the executed actions and allows the subsequent candidate linkage actions to obtain the optimal execution order based on the latest evidence information. This incremental reordering method has lower computational overhead than global re-solution and can quickly complete the reordering calculation in scenarios where feedback information arrives frequently, ensuring the real-time response capability of the system.

[0124] When the forced preemption condition is met, the system allows subsequent candidate linkage actions of high-risk event clusters to preempt the execution resources of low-risk event clusters. The forced preemption condition is that the updated true risk score of an event cluster exceeds a preset forced preemption threshold, and there are other event clusters with true risk scores lower than the updated true risk score of the event cluster that are currently occupying linkage resources. The preset forced preemption threshold is greater than a preset high threshold for true risk scores and less than or equal to 1. This value is configured by the system administrator according to the scarcity of linkage resources in the park and the security management strategy. In parks with abundant linkage resources, it can be appropriately increased to reduce unnecessary preemption, and in parks with scarce linkage resources, it can be appropriately decreased to ensure that high-risk events receive priority resource protection. When the forced preemption condition is met, the execution candidate linkage actions of low-risk event clusters are suspended, and the released linkage resources are allocated to the execution candidate linkage actions of high-risk event clusters.

[0125] The updated candidate action sequences of each event cluster are re-aggregated at the park level to obtain the updated total execution queue, and subsequent candidate actions are executed according to the updated total execution queue; thus, the order control during execution is reordered, rather than being fixed once triggered.

[0126] The reason for introducing the forced preemption mechanism is that in a park fire safety scenario, multiple event clusters with different risk levels may appear successively in a short period of time. When linkage resources are limited, the low-risk event clusters that arrive first may have already occupied some linkage resources, causing the high-risk event clusters that arrive later to be unable to obtain sufficient resources for timely response. If the system strictly allocates linkage resources according to the principle of first-come, first-served, the response time of high-risk event clusters will be blocked by low-risk event clusters, which may lead to serious safety consequences. The forced preemption mechanism sets a preset forced preemption threshold, allowing high-risk event clusters with a true risk score exceeding the threshold to suspend the pending candidate linkage actions of low-risk event clusters and take over the linkage resources they occupy, ensuring that high-risk events can always obtain priority resource guarantees when resources are scarce. The preset forced preemption threshold is set to a value greater than the preset high threshold of true risk score, ensuring that only event clusters with extremely high risk levels can trigger preemption, avoiding frequent resource preemption that leads to instability in linkage execution.

[0127] If new feedback information is continuously received within the preset listening time, the incremental update and partial rearrangement process in this step is repeated until the event cluster completes the processing loop or the preset listening time expires; when the event cluster completes the processing loop, proceed to step 500.

[0128] Step 500: After the event cluster completes the closed-loop processing, the warning parameters are reversed based on the closed-loop conclusion, and warning information is generated based on the warning parameters.

[0129] After the event cluster completes the closed-loop processing, the complete processing information of the event cluster is recorded; the closed-loop processing means that all candidate linkage actions corresponding to the event cluster have been executed or canceled, and a closed-loop conclusion has been obtained.

[0130] The purpose of this step is to use the closed-loop results of event cluster handling to reverse-correct the system's early warning parameters and generate proactive early warning information for areas that may be affected by related events, thereby constructing a complete adaptive loop from event access to closed-loop correction to proactive early warning. Traditional fire linkage systems usually treat the handling of each event as an independent process and do not update system parameters after handling, resulting in the system's inability to learn and improve from historical handling experience. This step feeds the closed-loop conclusions into early warning parameters such as historical false alarm rate, node health baseline parameters, regional basic risk coefficient, and resource occupancy cost benchmark value, enabling the system to continuously optimize its risk assessment and linkage decision-making capabilities based on actual handling results. At the same time, by generating early warning information for adjacent areas of closed-loop event clusters, the system can implement proactive prevention in potentially affected areas before the risk is completely eliminated, transforming passive response into proactive prevention.

[0131] The handling information includes the closed-loop conclusion, the execution chain of candidate linkage actions, the actual time consumed, the video hit status, the wireless link quality table, and the resource usage results. Among them, the closed-loop conclusion of the event cluster is whether the event cluster is ultimately confirmed as a real risk or a false alarm; the execution chain of candidate linkage actions is the complete execution sequence of the event cluster from the first batch of candidate linkage actions to the final candidate linkage actions and the execution results of each candidate linkage action; the actual time consumed is the total time from the generation of the event cluster to the handling closure; the video hit status is the consistency between the video review results and the closed-loop conclusion; the wireless link quality performance is the change of network quality parameters of each wireless node in the event cluster during the handling process; and the resource usage results are the types and durations of linkage resources actually used during the handling of the event cluster.

[0132] The system reverse-corrects the warning parameters based on the information recorded in the closed loop, which is used for the adaptive sorting of subsequent events. The warning parameters include the historical false alarm rate, node health, regional basic risk coefficient, and resource consumption cost benchmark value.

[0133] Among them, the health baseline parameter refers to the long-term communication reliability benchmark value that is pre-maintained and continuously updated for a single wireless node, used to characterize the basic health level formed by the wireless node during its historical operation. This parameter is not equivalent to the single-node health value calculated when a single event is reported, but is formed by smoothing and updating the actual single-node health value during multiple handling processes, and is used to reflect the long-term reliability of the wireless node in subsequent node health calculations and evidence consistency score calculations. The regional basic risk coefficient refers to the inherent risk benchmark value that is pre-maintained for a fire compartment and can be corrected according to the closed-loop conclusion, used to characterize the background fire risk level of the fire compartment when the current event has not occurred. This parameter is jointly determined by the functional attributes of the fire compartment, the distribution of combustibles or hazardous materials, historical fire or false alarm records, population density, evacuation conditions, and fire protection facility configuration, and is used as the regional risk background input in subsequent calculations of real risk score and early warning level score.

[0134] The reason for selecting historical false alarm rate, node health, regional basic risk coefficient, and resource consumption cost benchmark value as the target early warning parameters for reverse correction is that these four parameters correspond to the key input quantities for evidence consistency score calculation in step 200 and action priority value calculation in step 300, respectively. Correction of these four parameters can be transmitted to the final execution order of linkage actions through the score calculation link. Correction of historical false alarm rate gradually reduces the evidence contribution of wireless nodes that frequently falsely report incidents in subsequent events, and gradually increases the evidence contribution of wireless nodes that are frequently confirmed as real risks in subsequent events. Correction of node health baseline parameter gives wireless nodes with poor communication quality in the long term a lower credibility weight in subsequent events. Correction of regional basic risk coefficient gives areas that have experienced multiple real fires in history a higher risk assessment benchmark in subsequent events, and areas that have been repeatedly confirmed as false alarms in history a lower risk assessment benchmark in subsequent events. Correction of resource consumption cost benchmark value makes the estimated resource consumption of various linkage actions gradually approach the actual consumption level, improving the accuracy of the resource consumption cost item in action priority value calculation.

[0135] For each wireless node in the event cluster, its historical false alarm rate is updated based on the closed-loop conclusion. If the closed-loop conclusion is a false alarm, the historical false alarm rate of the wireless node increases; if the closed-loop conclusion is a real risk, the historical false alarm rate of the wireless node decreases. The update method is to append the closed-loop conclusion of the current event and its corresponding event timestamp to the tail of the event record queue of the wireless node, and at the same time check whether there are expired records at the head of the queue that have exceeded the rolling time window corresponding to the preset historical period. If so, they are removed from the queue. After the update, the historical false alarm rate of the wireless node is equal to the number of event records in the current queue that are confirmed as false alarms divided by the total number of event records in the current queue. In this way, the update of the historical false alarm rate is consistent with the calculation method based on the rolling time window set in step 200. Expired samples are automatically eliminated in each update to ensure that the historical false alarm rate always reflects the actual false alarm level within the preset historical period.

[0136] Based on the wireless link quality performance of each wireless node during the event cluster handling process, the health baseline parameters of each wireless node are updated to reduce the node health calculation results of wireless nodes with poor communication quality in subsequent events. Specifically, the system uses a preset health decay factor to update the health baseline parameters of the wireless nodes using an exponentially weighted moving average. The preset health decay factor has a value range of greater than 0 and less than 1. This value is configured by the system administrator according to the expected adjustment speed of node health. A larger value makes the health baseline more sensitive to the most recent handling result, while a smaller value makes the health baseline smoother and more stable. The update method is to multiply the original health baseline parameter of the wireless node by 1 and subtract the preset health decay factor, and then add the difference to the product of the actual single-node health value of the wireless node during the current handling process and the preset health decay factor. The result is the updated health baseline parameter.

[0137] The reason for using an exponentially weighted moving average to update the baseline health parameters of wireless nodes is that this method can gradually incorporate the latest health observations while retaining historical health information, resulting in a smooth and gradual trend in the changes of the health baseline parameters rather than drastic jumps. The exponentially weighted moving average method controls the fusion ratio of old and new information through a preset health decay factor. A smaller preset health decay factor allows historical information to dominate, making the changes in the health baseline parameters smoother and more stable, and filtering out accidental fluctuations in a single processing process. A larger preset health decay factor gives the latest observations a greater weight, allowing the health baseline parameters to track the actual changing trend of wireless node communication quality more quickly. This adaptive update method does not require storing all historical health observations; it only needs to maintain a health baseline parameter to complete the update calculation. The storage and computational costs are both in the constant range, making it suitable for large-scale campus environments with a large number of wireless nodes.

[0138] Based on the closed-loop conclusion of the event cluster, the regional basic risk coefficient of the fire compartment where the event cluster is located is updated. If the area repeatedly experiences event clusters that are confirmed as real risks, the regional basic risk coefficient of the area is gradually increased. If the area repeatedly experiences event clusters that are confirmed as false alarms, the regional basic risk coefficient of the area is gradually decreased. Specifically, the system uses a preset regional risk adjustment step size to update the regional basic risk coefficient. The preset regional risk adjustment step size is a positive real number greater than 0, and this value is configured by the system administrator according to the expected adjustment sensitivity of the regional risk level. When the closed-loop conclusion is a real risk, the regional basic risk coefficient of the area is added to the preset regional risk adjustment step size. If the result is greater than 1, it is set to 1. When the closed-loop conclusion is a false alarm, the regional basic risk coefficient of the area is subtracted from the preset regional risk adjustment step size. If the result is less than 0, it is set to 0.

[0139] Based on the execution chain and resource usage results of candidate linked actions in the event cluster, the resource usage cost of various candidate linked actions is updated so that the deviation between the actual resource consumption and the estimated value is corrected in subsequent calculations. Specifically, the system uses a preset cost correction learning rate to correct the preset resource usage cost benchmark value of various candidate linked actions. The preset cost correction learning rate is greater than 0 and less than 1, and this value is configured by the system administrator according to the expected convergence speed of the cost parameters. The update method is to add the original preset resource usage cost benchmark value of the candidate linked action to the product of the preset cost correction learning rate and the difference between the normalized value of actual resource consumption and the original preset resource usage cost benchmark value. The result is the updated preset resource usage cost benchmark value.

[0140] The reason for adopting an incremental correction method based on the learning rate to update the baseline value of resource occupancy cost is that the actual resource consumption of linked actions is affected by a variety of random factors. The actual resource consumption in a single process may deviate from the normal level due to special circumstances. If the actual resource consumption value is directly replaced with the new baseline value, the abnormal resource consumption in a single instance will cause the baseline value to jump drastically, affecting the stability of the priority value calculation for subsequent actions. By introducing a preset cost correction learning rate to control the magnitude of each correction, the system only adjusts the baseline value towards the actual consumption value by a limited step after each loop closure, so that the baseline value gradually converges to the statistical mean level of the actual resource consumption after multiple loop closures. This gradual correction method can adapt to the long-term changing trend of linked resource consumption patterns and filter out the accidental fluctuations in a single process, ensuring the stability and accuracy of the baseline value of resource occupancy cost.

[0141] After completing the reverse correction of the warning parameters, warning information is generated for areas within the park that may be affected by the related events, based on the handling information of the current event cluster and the updated warning parameters. The warning information refers to a proactive risk alert issued to areas or wireless nodes that have potential risks but have not yet experienced an event, based on the handling results of the closed-loop event cluster. This alert is used to deploy preventative measures or increase the level of attention for relevant areas before the next round of events arrives. The warning information includes a warning identifier, a warning generation timestamp, the identifier of the closed-loop event cluster that triggered the warning, the effective duration of the warning, and the warning content. The warning identifier is a unique code assigned to each warning message by the system; the warning level score is a quantitative value of the potential risk level faced by the warning target area at the current moment; the effective duration of the warning is the effective duration of the warning message; and the suggested preventative measures are response plans matched by the system from a preset preventative measures library based on the functional attributes of the warning target area. The specific values ​​of each field in the warning information vary depending on the warning level. High-level warnings include all of the above fields, while medium- and low-level warnings include subsets of the above fields. Specific differences are detailed in the warning content output section.

[0142] The process of generating the early warning information includes three stages: identifying the early warning object, calculating the early warning level, and outputting the early warning content.

[0143] The reason for dividing the early warning information generation process into three stages—early warning object identification, early warning level calculation, and early warning content output—is that early warning information generation involves three logically independent but data-dependent processing stages: spatial range determination, risk level quantification, and response measure matching. The early warning object identification stage determines which areas and wireless nodes need to receive early warning information, providing the spatial range for subsequent stages. The early warning level calculation stage quantifies the potential risk level of each early warning target area, providing the intensity basis for subsequent stages. The early warning content output stage generates differentiated early warning content based on the early warning level and sends it to the corresponding terminals, completing the entire process from early warning information generation to delivery. The layered design of these three stages ensures that the processing logic of each stage is relatively independent, facilitating independent configuration and optimization of parameters for different stages by system administrators. It also facilitates upgrading or replacing the algorithm of a single stage without changing the overall early warning process.

[0144] In the early warning target identification stage, based on the closed-loop conclusion and the area identifier of the closed-loop event cluster, the target area and target wireless node that need to receive the early warning information are determined. When the closed-loop conclusion of the closed-loop event cluster is a real risk, all adjacent fire zones in the park's spatial mapping information of the fire zone where the event cluster is located are determined as early warning target areas. The method for determining adjacent fire zones is consistent with the adjacent definition used in the spatial proximity function in step 100, that is, two areas share a physical boundary or belong to different floors of the same building unit in the park's spatial mapping information. At the same time, the building units to which each wireless node in the closed-loop event cluster belongs are checked, and those within the same building unit that have not yet been included in any currently active events are identified. Other fire zones within the cluster are also included in the early warning target area; all wireless nodes deployed within the early warning target area are marked as early warning target nodes; when the closed-loop conclusion of a closed-loop event cluster is a false alarm, no early warning information is generated for adjacent areas, but an equipment maintenance early warning is generated for nodes in the fire zone where the event cluster is located whose historical false alarm rate exceeds a preset high false alarm rate threshold; the preset high false alarm rate threshold ranges from greater than 0 to less than 1, and this value is configured by the system administrator according to the park's tolerance for false alarms. In parks with low tolerance for false alarms, it can be reduced to expand the coverage of equipment maintenance early warnings, and in older parks where the frequency of false alarms is generally high, it can be increased to avoid generating too many maintenance early warnings.

[0145] In the early warning level calculation process, an early warning level score is calculated for each early warning target area. This score quantifies the potential risk level faced by the target area at the current moment, ranging from 0 to 1, with higher values ​​indicating higher potential risks. The early warning level score is calculated by summing the following three products: the actual risk score of the closed-loop event cluster multiplied by a preset early warning risk transmission coefficient; the regional basic risk coefficient of the target area itself multiplied by a preset early warning regional weight coefficient; and the spatial proximity function value between the target area and the fire compartment containing the closed-loop event cluster multiplied by a preset early warning spatial weight coefficient. The preset early warning risk transmission coefficient is greater than 0 and less than or equal to 1, representing the attenuation of the actual risk transmission from the closed-loop event cluster to adjacent areas. This value is configured by the system administrator based on the fire isolation level of the park's building structure. In parks with comprehensive fire isolation measures, it can be reduced to weaken the risk transmission effect; in parks with strong building connectivity, it can be increased to enhance the risk transmission effect. The preset early warning regional weight coefficient... Both the coefficient and the preset warning space weight coefficient are greater than 0 and less than 1. The sum of the weight coefficients corresponding to the implicit weight coefficient, the preset warning area weight coefficient, and the preset warning space weight coefficient corresponding to the product of the preset warning risk transmission coefficient and the actual risk score of the closed-loop event cluster is equal to 1, and each weight coefficient is greater than 0. The specific values ​​are configured by the system administrator according to the park's warning strategy. When adjusting, it is necessary to ensure that the sum of the three weight coefficients is always equal to 1. When the warning level score is greater than or equal to the preset high warning threshold, the warning level of the warning target area is determined to be a high-level warning. When the warning level score is greater than or equal to the preset low warning threshold and less than the preset high warning threshold, the warning level of the warning target area is determined to be a medium-level warning. When the warning level score is less than the preset low warning threshold, the warning level of the warning target area is determined to be a low-level warning. The value range of the preset high warning threshold is greater than the preset low warning threshold and less than or equal to 1. The value range of the preset low warning threshold is greater than 0 and less than the preset high warning threshold. The above two thresholds are configured by the system administrator according to the park's warning sensitivity requirements.

[0146] In the early warning content output stage, differentiated early warning content is generated based on the early warning level and distributed to relevant personnel message push terminals and on-site handling terminals through the park's fire control platform. For high-level early warning target areas, the generated early warning content includes the fire zone code of the early warning target area, the early warning level score, the area identifier and closure conclusion of the closed-loop event cluster that triggered this early warning, the node identifiers and current node health of all early warning target nodes within the early warning target area, and recommended preventive measures. The recommended preventive measures are obtained by the system from a preset preventive measures library based on the functional attributes of the early warning target area. The preset preventive measures library is pre-configured by the system administrator and includes, but is not limited to, increasing the frequency of inspections, pre-positioning fire extinguishers, notifying the area's responsible person to increase vigilance, and activating video surveillance for continuous monitoring. At the same time, the monitoring sensitivity of all early warning target nodes within the high-level early warning target area is temporarily increased. Specifically, the alarm threshold of the early warning target node is temporarily reduced by a preset early warning sensitivity increase ratio, the value of which is greater than a certain value. The value is 0 and less than 0.5. This value is configured by the system administrator based on the expected detection sensitivity during the warning period. It can be appropriately increased in scenarios requiring high-sensitivity detection and appropriately decreased in wireless node environments with high false alarm rates to avoid excessive false alarm events during the warning period. The temporary reduction of the alarm threshold will automatically restore to the original value after the preset warning validity period expires. The preset warning validity period is a positive real number greater than 0. This value is configured by the system administrator based on the park's fire inspection cycle and the duration of the risk. It is shortened in parks with shorter inspection cycles and extended in scenarios where the risk may last for a longer period of time. For warning target areas with medium-level warnings, the generated warning content includes the fire compartment code of the warning target area, the warning level score, and the area identifier of the closed-loop event cluster that triggered this warning. The system pushes the warning notification to the person in charge of the area but does not adjust the monitoring sensitivity of the warning target node. For warning target areas with low-level warnings, the system only records the warning information in the warning log of the park's fire control platform and does not actively send warning notifications to personnel message push terminals.

[0147] For equipment maintenance early warnings, the system will generate an equipment maintenance early warning report by summarizing the node identifier, area identifier, current historical false alarm rate, and current node health status of wireless nodes whose historical false alarm rate exceeds the preset high false alarm rate threshold. The report will be sent to the message push terminals of system administrators and equipment maintenance personnel through the park's fire control platform. The equipment maintenance early warning report will also include the time distribution of false alarm events and the trend of network quality parameter changes for the wireless node within a preset historical period to help maintenance personnel determine the root cause of node anomalies and formulate maintenance plans.

[0148] The system assigns a unique warning identifier to each warning message and records the warning identifier, warning target area, warning level score, warning generation timestamp, closed-loop event cluster identifier that triggered the warning, and warning validity period in the warning information storage queue. When the warning validity period expires, the system marks the warning message as expired and restores the node monitoring sensitivity parameters that were temporarily adjusted during the warning period. If a new event cluster is generated in the warning target area within the warning validity period, the system uses this warning message as an auxiliary reference during the event merging process in step 100, assigning a correlation coefficient to events from the warning target area when calculating the correlation of the new event cluster. A preset warning correlation bonus value is assigned; this preset warning correlation bonus value is a small positive value, which is configured by the system administrator based on the expected impact of the warning information on the subsequent event correlation judgment. The value range is limited to a small positive range to ensure that the bonus mechanism has a moderate and auxiliary effect on the correlation calculation rather than a dominant and mandatory intervention. Through the above warning correlation bonus mechanism, events from the warning target area within the warning validity period receive a moderate increase in correlation in the correlation calculation, enabling the system to maintain a higher level of attention to new events within the warning area, thereby achieving proactive prevention before the risk is completely eliminated.

[0149] The reason for introducing the early warning correlation bonus mechanism is that when a real fire has been confirmed in a certain area and the response loop has been closed, the potential risks faced by its adjacent areas in the short term are higher than normal. This is because the spread of the fire, the diffusion of smoke, and the chain reaction of building structures may cause new abnormal events to occur in adjacent areas for a period of time after the loop is closed. If the system still processes new events in the early warning target area according to the conventional correlation calculation method, these new events may not be able to be merged into high-priority event clusters in a timely manner due to insufficient correlation, resulting in a delay in the system's response to potential risks. By assigning an appropriate bonus value to events from the early warning target area in the correlation calculation, the system can give higher attention to new events in the early warning area without changing the overall framework of the correlation calculation, making these events easier to be merged into event clusters and obtain higher risk assessment results, thereby achieving proactive prevention of potential risks. The preset early warning correlation bonus value is limited to a small positive range to ensure that the impact of the bonus mechanism on the correlation calculation is an auxiliary and moderate enhancement rather than a dominant and mandatory intervention, avoiding the over-merging of irrelevant events in the early warning area due to excessive bonus values.

[0150] Through the aforementioned reverse correction and early warning information generation, adaptive adjustment of early warning parameters and proactive risk warning are achieved.

[0151] The embodiments of this application have been described above, but these embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments based on the guidance of these embodiments, and all of them are within the protection scope of these embodiments.

Claims

1. A method for controlling the execution sequence of fire-fighting linkage in a park, characterized in that, include: Within a preset sliding time window, events reported by the wireless network within the park are received, and the events are converted into unified event vectors. Based on the correlation between events, the unified event vectors are merged into one or more event clusters. For each event cluster, multidimensional features are extracted, and an evidence consistency score is calculated based on the multidimensional features. The true risk score is then calculated by combining the preset regional basic risk coefficient. Based on the evidence consistency score and the true risk score, a set of candidate linkage actions is generated for each event cluster. An action execution sequence diagram is constructed based on the set of candidate linkage actions, and the priority execution sequence is solved. The total execution queue is obtained by summarizing the priority execution sequence based on the action execution sequence diagram. The first batch of candidate linkage actions are output and issued according to the total execution queue. Within the preset listening time, the evidence consistency score and the real risk score are updated according to the feedback information. The remaining candidate linkage actions are partially rearranged to obtain the updated total execution queue. Subsequent candidate linkage actions are executed according to the updated total execution queue. After the event cluster completes the closed-loop processing, the warning parameters are corrected in reverse based on the closed-loop conclusion, and warning information is generated based on the warning parameters.

2. The method for controlling the execution sequence of fire-fighting linkage in a park according to claim 1, characterized in that, Methods for constructing event clusters include: A time proximity function is set based on the event timestamps in the unified event vector, where the event timestamps are clock-corrected times. Set a spatial proximity function based on the region identifier in the unified event vector; Set event category-related functions based on event categories in the unified event vector; Set the network feature similarity function based on the set of network quality parameters in the unified event vector; The correlation degree between any two unified event vectors is calculated. The correlation degree is obtained by weighted summation of the temporal proximity function value, spatial proximity function value, event category correlation function value, and network feature similarity function value. When the correlation is greater than or equal to the preset merging threshold, the corresponding events are grouped into the same event cluster, and the event cluster is constructed based on the transitive closure.

3. The method for controlling the execution sequence of fire-fighting linkage in a park according to claim 2, characterized in that, After the event cluster is constructed, it also includes intra-cluster consistency verification. When the event cluster does not meet the consistency verification, the events that do not meet the constraints are removed from the event cluster and re-merged to obtain an event cluster that passes the verification. The consistency check is based on one or more of the following constraints: maximum spatial span within the cluster, cluster centrality constraint, and minimum event category consistency constraint within the cluster.

4. The method for controlling the execution sequence of fire-fighting linkage in a park according to claim 2, characterized in that, Methods for obtaining the event timestamp include: Each wireless node carries its local timestamp when reporting an event. When the event is forwarded by the gateway node, the gateway node records the gateway reception timestamp of the event. Calculate the total clock offset of the wireless node based on the gateway's received timestamp and the node's local timestamp; The event timestamp is obtained by adding the node's local timestamp to the total clock offset.

5. The method for controlling the execution sequence of fire-fighting linkage in a park according to claim 1, characterized in that, Methods for calculating the evidence consistency score include: Multidimensional features include multi-node co-occurrence consistency, video verification support, historical false alarm rate, and node health. The evidence consistency score is the weighted sum of node co-occurrence consistency, video verification support, 1 minus historical false alarm rate, and node health.

6. The method for controlling the execution sequence of fire-fighting linkage in a park according to claim 5, characterized in that, The methods for calculating the health of a node include: Calculate the retransmission health component based on the number of retransmissions by the wireless node. The signal strength of the wireless node is linearly normalized and mapped to the interval between 0 and 1 to obtain the signal health component; For each wireless node in the event cluster, the retransmission health component is multiplied by the signal health component to obtain the single node health value. The average of all single node health values ​​is then used to obtain the node health of the event cluster.

7. The method for controlling the execution sequence of fire-fighting linkage in a park according to claim 1, characterized in that, Methods for calculating the true risk score include: The sensor readings of all wireless nodes in the event cluster are normalized to obtain the event intensity of all events in the event cluster, and the maximum value of the event intensity is used as the event intensity aggregate value. The true risk score is the result of a weighted sum of the evidence consistency score, the regional basic risk coefficient, the preset impact coefficient of key targets, and the event intensity aggregate value.

8. The method for controlling the execution sequence of fire-fighting linkage in a park according to claim 1, characterized in that, Methods for generating a set of candidate action linkages include: The candidate linkage action set consists of several candidate linkage actions corresponding to the event cluster; When both the true risk score and the evidence consistency score are greater than or equal to their respective preset high thresholds, a set of candidate linkage actions, including review, notification and escalation actions, is generated. When the true risk score is greater than or equal to the preset high threshold of true risk score, but the evidence consistency score is less than the preset high threshold of evidence consistency score, or when the true risk score is between the preset high threshold of true risk score and the preset low threshold of true risk score, a set of candidate linkage actions including review and notification actions is generated. When the actual risk score is less than the preset low threshold for the actual risk score, a set of candidate linkage actions, including review actions, is generated.

9. The method for controlling the execution sequence of fire-fighting linkage in a park according to claim 1, characterized in that, Methods for obtaining the total execution queue include: The action execution sequence graph is a graph with candidate linked actions as nodes and directed edges with the dependencies between candidate linked actions; The time benefit of the candidate linkage action is calculated based on the preset timeliness benchmark value of the candidate linkage action and the duration of the event cluster. The resource cost of the candidate linkage action is calculated based on the preset resource cost benchmark value of the candidate linkage action. The action priority value of the candidate linkage action is calculated based on the actual risk score, time benefit, resource cost and preset evidence clarification gain. Based on action priority values, and under the premise of satisfying the dependencies between events and the mutual exclusion between candidate linked actions, the priority execution sequence is obtained by solving all candidate linked actions of the event cluster. The overall execution queue is obtained by summarizing the priority execution sequences based on the action execution sequence diagram.

10. The method for controlling the execution sequence of fire-fighting linkage in a park according to claim 1, characterized in that, The methods for obtaining the updated total execution queue include: After the first batch of candidate linkage actions are issued, a feedback listening point is set for each event cluster to listen for feedback information within a preset listening time. If the feedback monitoring point receives new feedback information within the preset monitoring time, the new feedback information is mapped as incremental evidence, and the evidence consistency score and true risk score of the event cluster are recalculated based on the incremental evidence. Based on the recalculated real risk score, the action priority value of each candidate linkage action that has not yet been executed is calculated. While keeping the dependencies and mutual exclusion relationships unchanged, the remaining candidate linkage action sequence is locally rearranged to obtain the updated candidate linkage action sequence. The updated candidate action sequences are re-aggregated to obtain the updated total execution queue.

11. The method for controlling the execution sequence of fire-fighting linkage in a park according to claim 1, characterized in that, Methods for generating early warning information include: When the closed-loop conclusion is a real risk, the adjacent fire compartments of the fire compartment where the event cluster is located and other fire compartments within the same building unit are identified as early warning target areas; The warning level score is calculated by weighting the actual risk score of the closed-loop event cluster, the regional basic risk coefficient and the spatial proximity function value in the warning parameters of the warning target area; The warning content is generated and issued based on the warning level score; Early warning information is generated based on the warning content.

Images